Saturday May 3
Time Name Title
8:00-8:15 Intro
8:15-8:35 Erhard Bruderer From The Margin To The Mainstream: An
Martin Maiers Agenda For Computer Simulations In The
Social Sciences
8:35-8:55 Roger M. Stein An Adaptive Simulation Approach for
Robert N. Bernard Investigating Information Processing
Structures in Organizations
8:55-9:15 Walid Nasrallah The Firm as Cocktail Party:
Organizational Effectiveness in a
Barrier-free World
9:15-9:30 Open Discussion
9:30-10:00 BREAK
10:00-10:20 Gabor Peli Market Partitioning and The Geometry of the
Bart Nooteboom Resource Space
10:20-10:40 Tomas B. Klos Managing dependence in vertical interfirm
relations
10:40-11:00 Jeroen Bruggeman Distribution of competition in
Ivar Vermeulen organizational markets.
11:00-11:20
11:20-11:35 Open Discussion
11:35-1:00 LUNCH
1:00-1:20 Anjali Sastry The road not taken: Simulating path
dependence in organizational evolution
1:20-1:40 Kathleen Carley Organizational Adaptation in a Volatile
JuSung Lee Environment
1:40-2:00 Bill McKelvey Kauffman's NK Model of Complexity as Cause:
A Critique
2:00-2:15 Open Discussion
2:15-2:45 BREAK
2:45-3:05 Timothy Van Zandt Real-Time Hierarchical Resource Allocation
3:05-3:25 Jan Thomsen Designing Quality into Product Development
John C. Kunz, Yul K. Kwon, Organizations Through Computational
Sam Miller, Raymond E. Levitt Organizational Modeling and Simulation
3:25-3:45 Oliver Guenther Market Organizations for Controlling Smart
Tad Hogg Matter
Bernardo Huberman
3:45-4:05 Open Discussion
4:05-5:00 Panel - Carley, Sastry, Issues in Modeling and Analyzing
McKelvey Organizations from a Complexity Perspective
Sunday May 4
Time Name Title
8:00-8:15 Intro
8:15-8:35 Jaap Kamps Computational Support for Logical
Formalization
8:35-8:55 Corrado Pasquali PI-calculus and the Logic of Organizations
8:55-9:15 Laszlo Polos Challenges for Organizational Learning:
l Issues in Logical Modelling
9:15-9:30 Open Discussion
9:30-10:00 BREAK
10:00-10:20 Lars Christensen Learning organizations in action - a system
Tore Christiansen dynamics model of satisficing search with
adaptive aspirations
10:20-10:40 Robert N. Bernard The Evolution of Skill Sets in Consulting
Firms
10:40-11:00 Carter Butts A Bayesian Model of Panic in Belief
11:00-11:20 Kevin Crowston A Simulation of a Network of Agents Solving
Bernardo Huberman Travelling Salesman's Problems
11:20-11:35 Open Discussion
11:35-1:00 LUNCH
1:00-1:20 Raymond E. Levitt Promoting or Discouraging Conflict in
Jan Thomsen, Yul K. Kwon Engineering Project Teams: Insights from a
Computational Organizational Design
Perspective
1:20-1:40 Douglas B. Fridsma, M.D. Modeling Medical Processes for
Jan Thomsen Computational Organization Simulation:
Linking Protocol Descriptions to Simulation
Requirement
1:40-2:00 Scott Moss, SDML: A Multi-Agent Language for
Helen Gaylard, Steve Wallis, Organizational Modelling
Bruce Edmonds
2:00-2:15 Open Discussion
2:15-2:45 BREAK
2:45-3:05 William E. Paterson (Bill) A Garbage Can Simulation Model of High-Risk
Facility Siting
3:05-3:25 Daniel E. O'Leary Virtual Organizations
3:25-3:45 Yan Jin PIF
3:45-4:05 Open Discussion
4:05-4:10 Announcements
4:10-5:00 Panel - Bruderer, Levitt, Linking Data and Models
O'Leary
FROM THE MARGIN TO THE MAINSTREAM: AN AGENDA FOR COMPUTER SIMULATIONS IN THE SOCIAL SCIENCES
Erhard Bruderer
The University of Minnesota
Carlson School of Management
and
Martin Maiers
The University of Minnesota
Department of Computer Science
Computer simulation has had a long and established tradition in the social sciences. Unfortunately, research in this field has always been at the margin. Why? After all, computer simulations allow us to make rigorous models of complex social processes which are not accessible to mathematical modeling. This paper explores why we failed to have a major impact on the social sciences. However, there is good news on the horizon which is likely to change our fate. Among the most important changes are the development of the Internet and the emerging standard of the computer language Java. We argue that in the future the Internet and Java will make computer simulation easily accessible to many social scientists and thus will spur high growth in its use. In this paper we outline a specific and practical agenda to move computer simulation from the margin into the mainstream of social science research. We conclude that 30 years from now, computer simulation will be a mainstream research tool in the social sciences as much as advanced statistics is today.
AN ADAPTIVE SIMULATION APPROACH FOR INVESTIGATING INFORMATION PROCESSING STRUCTURES IN ORGANIZATIONS
Roger M. Stein
Leonard N. Stern School of Bus
New York University and Moody's Investors Service
and
Robert N. Bernard
Department of Urban Planning
Rutgers University and Coopers & Lybrand Consulting
We adopt the view of organizations as information processing entities. We present a methodology for modeling organizational structures and for determining which organizational structures, if any, distinguish themselves given various constraints. As such, we propose that various exogenous and endogenous factors should affect the performance of organizations with respect to information processing tasks. To test our propositions we implement a simulation methodolgy. Our methodology relies heavily upon combining Monte Carlo methods and genetic algorithms to represent dynamic organizational operating environments and competition among firms, respectively. We present examples of the organizational computer model and demonstrate results of the methodology for testing our underlying propositions.
We feel the contribution of our work is not so much the valdiation or rejection of specific propositions, but rather the introduction of a new tool for validating entire classes of organizational prosositions through simulation. The reimainder of this extended abstract foucuses on an overview of this methodolgy.
Organizations presumably do not know the functional form of their information processing dynamics. As such, we postulate that organizations perform a type of trial and error search for effective structures. Simulation methods allow us to mimic this search process and test the propositions while making only a limited number of assumptions about the parameters of the problem. They also allow us to deal with the complexity inherent in organizational structures. Finally, simulation allows us to create an environment in which differing organizational structures compete with each other to efficiently perform similar tasks.
We have created a generic toolbox, called OrgNet, for representing organizations. The toolbox is an object library written in C++. The toolbox allows us to represent organizations, their members, and their structures. The object library represents organizations' members as having a fixed amount of memory, a particular computing speed, a task function, and a message queue that holds input information that the member receives. The organization itself is composed of a number of members and a matrix which defines the various types of relationships among the members.
The computing speed parameter indicates how many units of time the member requires to execute its task function. The task function itself defines how the member processes the information in its message queue. The message queue serves as a post-office or clearing house for information. Members communicate with each other by posting an unambiguously addressed message to the message queue. The organization coordinates delivery of these messages to its members. Finally, the organization has a clock which synchronizes the activity of the organization members and message delivery. Each message is time-stamped with its scheduled delivery time and is delivered at the time indicated.
Organizational tasks can take on a variety of forms. For simplicity, we adopt a definition of tasks as either (1) linear decision rules, or (2) pattern matching. In the first case, the goal of the organization is the make decisions based on a linear combination of environmental inputs. In the second, the decision makers in the organization must compare environmental input with a set of known exemplar patterns and make decisions based on the distance between them.
Once an organization has been defined, we create an environment in which it can operate. In order to create such an environment, we perform stochastic (Monte Carlo) simulation. Using random number generators, we produce a time-series of inputs (environmental cues) which the simulated organization must process.
The time-series is processed by the members as follows:
1. At each tick of the organization clock, the time-series is evaluated.
2. If an input is not present in the time-series, no action is taken.
3. If an input is present it is placed into the organization queue addressed to all organization members eligible to receive input from the environment.
4. The input is delivered to the queue of individual organization members who process it according to their task functions.
Only the timing of the inputs is random, the actual form of the inputs is a simple periodic function, either additive or pattern-matching.
To simulate inter-organizational competition we invoke a genetic algorithm-based approach. By comparing the actual output of an organization with the ideal output that the organization would have gotten if it processed the information "perfectly," we can get a feel for how well the organization processes the input information from the environment.
Since the input data is given and represent functions that are deterministic, we know the ideal outputs the organization should produce. The difference between the actual output of an organization and the desired output form the error of an organization in performing its tasks. The larger the error, the poorer the organization's performance.
By comparing any two organizations, we can determine which is more efficient than the other by comparing the error rates of the two. The organization with the lower error rate is de facto more efficient at solving the particular problem at hand in that particular environment. Note that this does not necessarily mean that the favored organization is most efficient at solving all other problems. In fact, it often means that it will not be.
Since we represent an organization as an individual chromosome, our GA can create a population of organizations, each with a different structure. The GA then allows these organizational structures to compete with each other at solving a task (in our case, the task represented by processing the time series generated by the Monte Carlo simulation). By periodically evaluating the performance of each organization at the task (i.e., the accuracy of a linear decision or the correct classification rate for a pattern matching problem) the GA is able to rank each of the organizations based on its performance. This continues for several generations.
The idea of these simulations is to test the propositions by exploring the range of organization dynamics at different levels of the various environmental and organizational parameters (e.g., memory size, speed of information, computational speed). To do this we will perform a large number of runs of individual genetic algorithms. The output of each run of the genetic algorithm will be a "most fit" organizational structure, given the particular set of parameters. Of course, this requires that we simulate many structures at many different levels of each parameter.
The methodology we propose offers one means to testing empirically the effects of limitations on individual cognitive powers and the severity of information environment on information processing in various organizational structures. By testing the sensitivity of competing organizational structures to changes in environment or limitations on internal resources, we can determine, through simulation and simulated competition, which structures are efficient under which conditions.
Note: This work draws heavily on work done by Stein: "Does Organizational Structure Matter? An Adaptive Simulation Approach for Investigating Information Processing Structures in Organizations," Stein, R., Working Paper #IS-96-4, Information Systems Working Paper Series, Stern School of Business, 1996.
THE FIRM AS COCKTAIL PARTY: ORGANIZATIONAL EFFECTIVENESS IN A BARRIER-FREE WORLD
Walid Nasrallah
Stanford University
Construction Engineering & Management Division
Department of Civil Engineering
Time was when communication was costly. It took the resources of largish empire to afford the infrastructure for regular communication, and that was only justified by national defense. In the world of commerce, again only the very largest of mercantile concerns had use for chains of watch towers and carrier pigeon routes. Today, by almost incomprehensible contrast, the pace of quantitative and qualitative improvements in the means of communication has accelerated to such a frenzy that it hardly pays to try and gauge the effects on (and hence value to) an enterprise of adopting newly invented communication systems: the new systems become so cheap so rapidly that they get adopted by default before the evaluation process has had time to be completed.
A researcher who wants to predict the trajectory of future organizational evolution is better advised to look at the eventual steady state of this process of technological advance: namely, a hypothetical but no longer fanciful state where communicationband width goes to infinity and cost to zero. At that point, only human information processing capacity and relative utility of communicating with various individuals in an organization Will matter when choosing whom to interact with and how often. The corporation of the future (e.g. Zuboff "In the Age of the Smart Machine" [1] and Davidow's "The Virtual Corporation"[2]) have the same free-form, mingling structure as a stereotypical cocktail party.
This paper attempts to build a rigorous framework for deriving communication patterns from basic measures of communication utility. I build upon the assumptions of Huberman and Hogg in their "Communities of Practice" paper [3], which investigated the genesis and size of groups within a population whose members engage in sporadic, one-way, pair-wise interactions. The basic premise of therefore mentioned paper was that when an individual in the population seeks to interact with another individual (let us call the latter the "source" and the former the "seeker"), then this interaction will only be valuable to the seeker if it is relatively unprecedented, i.e., if sufficient time has elapsed since the last interaction with the same source.
The effect of this assumption is to make it less attractive to just choose a favorite source and interact only with him or her. In addition, the time needed for a source to once again become useful(to a particular seeker) was assumed to be randomly distributed. This means that no seeker can uniquely specify a fixed subset of top sources with whom to interact if the goal is to maximize the sum of all interaction utilities. The model therefore expressed expected size of interacting cliques at the optimal global utility point, as a function of pair-wise utility variation and population size. The paper also set up the mathematics to investigate how a community might move towards this optimum given limited, inaccurate or outdated information. The authors presented their results as applicable to "Communities of Practice": groups of researchers, bureaucrats, or other professionals who share a common field.
To use the same framework for studying an organization where interactions are a staple of everyday work, we need to consider an additional limitation that prevents each individual from simply latching on to the most rewarding relationships. In this paper, I posit that an individual who becomes the target of many requests for interaction can only handle a limited number of them. In other words, the most useful people in the organization will have a queue at their door which reduces their overall usefulness as a source.
Adding this assumption to the Huberman-Hogg model makes it possible to ask many interesting questions. For example, we might: - find the best mix of generalists and specialists in the organization,
- gauge the cost of nagging "bad apples" who use up the time of expert sources who might be of more use to other seekers, or
- investigate the ramifications of alternative criteria (e.g. random, tit-for-tat, expectation future reciprocity) for selecting which seek to help first when a source has a queue outside his or her door.
The uses of this sort of model are open-ended. Beside the versatility of expressing specific cause-effect relationships as algebraically manipulable mathematical expressions, each algebraic form thus derived might have multiple meanings based on how real life observations were abstracted to provide the independent variables of the expression. For example, if we can measure directly the interaction structure and efficiency of an organization, we can use a particular result of the model to deduce the variance in relative pair-wise utilities. The interesting trick is that this same kind of model result can just as readily be used to predict the effects of a change in collaboration policy on overall productivity given the same mix of people.
Another strength of the model is that many different external effects, like natural and institutional environment or employee skill mix or task dependencies, can be distilled into representative modifications to the pair-wise utility matrix prior to entering the actual model. This gives an investigator a fair amount of latitude both in setting up the questions to ask and in interpreting the results, very much in the spirit of Cohen, March & Olsen's "Garbage Can Model"[4] by, but without the opacity of simulation or the artifice of modeling essentially stochastic processes with deterministic routines.
In the presentation, I will set up a few example scenarios for the model and derive properties of the sort of structure the model would predict for each scenario. I will illustrate the meaning of these predictions using the example of a futuristic equal-access virtual corporate structure. I will then re-frame the same mathematical results in the context of several other broad areas of more immediate practical interest for future research. Specifically, I will focus translating the results into three other contexts:
1) The non-business party, club or community, where utility of interaction is a matter of personal preference, but similar capacity limits exist.
2) The highly structured organization, where choice of possible sources is limited by geographical or institutional structure, but some latitude is left that can be studied using the same model as above.
3) The project organization, where structure needs to shift in response to the different phases of the project, and a specific subset of interactions is required to succeed in order to proceed to the next phase.
For the last two domains of application, I regard the current model to be a necessary but not sufficient underpinning for a general "theory-creating" effort, in much the same way as study of frictionless motion is a prerequisite for formalizing an understanding of mechanics.
REFERENCES:
Shoshana Zuboff, 1988, "In the Age of the Smart Machine: The Future of Work and Power", Basic Books, New York.
William H. Davidow, 1992, "The Virtual Corportation: Structuring and Revitalizing the Corporation for the 21st century" Harper Business, New York.
Bernardo A. Huberman and Tad Hogg,1995, "Communities of Practice, " Computational and Mathematical Organization Theory, 1, (1).
Michael D. Cohen, James G. March and Johan P. Olsen, 1972,"A Garbage Can Model of Organizational Choice," Adminstrative Science Quarterly, March, 17,(1).
MARKET PARTITIONING AND THE GEOMETRY OF THE RESOURCE SPACE
Gabor Peli - Bart Nooteboom
University of Groningen
Faculty of Management & Organization
The presentation proposes a geometric explanation for the findings of organization ecology's resource partitioning theory (Carroll, 1985; Carroll and Hannan, 1995). We show that the proposed geometric approach supports the theory's original, scale economy advantage based explanation in two ways. First, the geometric model unveils robust additional effects that drive the events towards the perceived market histories. Second, the combination of the "geometric" and "scale economy based" layers of explanation helps to answer open questions in the theory. We briefly summarize the main arguments of the original theory, and then we sketch the main contributions of the geometric approach.
The resource partitioning theory investigates the underlying dynamics of markets populated by generalist and specialist organizations. Themarket is modelled as an n-dimensional resource space, where customer demand constitutes the key resource for organizations. Demand is distributed along several taste categories. Since tastes are characterized by n characteristics, they can be seen as points in an n-dimensional resource space. An organisation's niche is the zone in the resource space from where the organisation attracts customers. Assuming that each taste characteristic is of similar importance (taste descriptors are standardised), organisational niches are n-dimensional spheres - hyperspheres - in the resource space. Generalists offer products to customers on a broad range (wide niche, big sphere radius); specialists address specific tastes (narrow niche, small sphere radius).
The resource partitioning theory claims that an increasing concentration of big generalists opens resource pockets for small specialist organizations at the marginal market positions. In mature markets, generalists and surrounding small specialists can realize a peaceful coexistence. This outcome is achieved in the following manner. Originally generalists compete with each other for market dominance, specialists survival chances are poor because of scale economy effects. Those generalists that address the abundant spots in the resource space grow bigger and finally they outforce the smaller generalists from the market (medium size organization outflow). As competition chills out, the few very big survivors can occupy the best, central zones of the resource space. In the meanwhile, they increase their niche width consuming additional resources. However, niche breadth increase has some inherent limits (principle of allocation, Hannan and Freeman, 1989), so some less abundant parts of the markets (e.g., marginal tastes) are passed over to little specialists.
It is crucial in the original resource partitioning explanation that the resource distribution is inhomogenous (market centers exist). The proposed geometric explanation assumes even resource distribution. Thus, it serves as a background layer for the original explanation in the sense that resource inequalities can be added as perturbations to the flat distribution. The geometric explanation applies the mathematical results of a domain in geometry, the so called sphere packing problem (Conway and Sloane, 1988). The general sphere packing problem is the following: fill up the n-dimensional Euclidean space with n-dimensional (hyper)spheres of equal radius in the possible densest manner. The success of a sphere packing is measured by "packing density", the ratio of the volume occupied by the spheres to total space volume.
The geometric interpretation considers one extra explanatory variable: the number of spatial dimensions. Since customer demands become more elaborated with time, the number of taste descriptors (resource space dimensions) increases as markets get older. The n-dimensional hyperspheres that represent organizational niches have to fold out into n+1 dimensions after each dimensional change.
The question is that what happens to packing density as n increases? The pertaining mathematical results show that the density of even the best packings dramatically decreases with the increasing number of spatial dimensions. In socio- economic terms: the volume of leftout space around generalist organizations robustly increases, more and more little pockets open that offer niches for tiny specialists.
Two other mathematical notions are also introduced to refine the explanations. The first, the "kissing number", is the number of spheres that touch a chosen sphere in a given sphere packing. The second, the "thickness of a covering" informs about the minimally required overlap (now: niche overlap) if spheres do not only touch but completely cover the space. The change of these two parameters reveals some effects that facilitate, for example, market partition between generalist organizations of the same size. The combination of the geometric and scale economy based layers of explanations eliminates some shortcomings in each approach, giving a much detailed account for the resource partitioning phenomenon.
REFERENCES:
Carroll, G. R. 1985. Concentration and Specialization: Dynamics of Niche Width in Populations of Organizations. American Journal of Sociology, 90: 1262-1283.
Carroll, G. R. and M. T. Hannan, 1995. Organizations in Industry. Strategy, Structure, and Selection. Oxford University Press, Oxford, New York.
Conway, J. H and N. J. A. Sloane, 1988. Sphere Packings, Lattices and Groups. A Series of Comprehensive Studies in Mathematics 290 (Grundlehren der mathematischen Wissenschaften) Springer Verlag, New York, Berlin.
MANAGING DEPENDENCE IN VERTICAL INTERFIRM RELATIONS
Tomas B. Klos
Faculty of Management and Organization
University of Groningen
1. Introduction
Increasing competition forces firms to differentiate their products and focus all---or at least most---of their resources to carrying out their core activities and to outsource other activities. The paper investigates the increasing number of relations between buyers and suppliers of industrial goods that come into being as a consequence of this development. Nowadays, rather than as a necessary evil, such relations are often seen as a potential source of advantage, if both partners let the other utilize the respective value each has for the other. However, the outcome of a relation depends on {\it both} firms' contributions which means they have to manage the relation to actually realize the potential returns inherent in the relation. There is not only more to gain, but also more to loose if the relation does not unfold as planned.
Whether or not this happens is assumed to be uncertain to the agents because of their bounded rationality and their limited information with respect to what happens in the environment and with respect to their partner's behavior (Williamson, 1985). The paper describes computer simulations of the developmental processes of these vertical interfirm relations, as guided by---among other things---the purposive management of a relation by the agents involved. In particular, the paper compares the performance of organizations using different forms of management under different circumstances.
2. The simulation model
2.1 Buyer-supplier relations
In our model, there are different types of agents. A specific agent can be a buyer, a supplier or both. In our simplest model, there are only pure buyers and pure suppliers, although buyers are in turn sellers on a final market. However, we are not focusing on such one-at-a-time transactions involving producers and consumers (such as Albin & Foley, 1992; Epstein & Axtell, 1996; Tesfatsion, 1997 and Vriend, 1995) but rather on relations between producers and other producers---their industrial suppliers.
Buyers experience a certain demand on their final market---the one where they sell their products. This demand has 2 dimensions, volume and specialization. Volume speaks for itself and specialization is a measure of the product-quality demanded. (Quality is actually the degree to which a product's characteristics correspond to the preferences of customers on the dimensions along which the product is differentiated.) Volume and specialization are inversely related, allowing the generic strategies of `cost leadership' and `differentiation' (Porter, 1985) as the extremes of the continuum that runs from low cost/high volume to high cost/low volume. Higher specialization thus implies that the products are more unique; we will also assume the profit margin on these products to be higher because customers are willing to pay more for products more suited to their specific preferences.
As its core competence, a buyer produces and sells products that satisfy a certain demand on the final market. The inputs for this production (components) it can either produce itself, or request a supplier to (produce and) supply---the `make or buy' decision. In either case, producing the inputs requires investments to be made in assets (e.g. machines). Assets are specific to the input produced to the extent that the final product for which the inputs are used is specialized. For example, assets used for producing the inputs which are in turn used for making a highly specialized product, can not be used for producing other inputs.
The efficiency with which inputs and products are produced depends on the volume of production, as a representation of scale economies. A buyer that produces its own inputs reaches lower economies of scale and therefore produces at a higher cost than a supplier that produces inputs for more buyers. However, if a buyer requests a supplier to produce, the supplier must make the investments in specific assets. The supplier will only want to do that if it thinks its interaction history with the requesting buyer was `pleasant' enough (cf. Williamson's `atmosphere'). In general, different events in the development of a relation between two agents influence each of those agents' subjective interpretation of the other agent's `loyalty'.
2.2 Management of dependence
There are various stages in the evolution of a relation. A buyer has to decide whether it wants to make or buy inputs, which it can not determine apart from a supplier to buy the inputs from. In this decision, the buyer looks at different suppliers' efficiency (which it can find out by sending them requests for proposals or RFPs) as well as loyalty. A supplier's efficiency co-determines the potential returns inherent in the relation whereas its loyalty gives the buyer an indication of the probability that these returns will actually be realized. The potential returns are also determined by the level of specialization of the products that the buyer intends to produce using the inputs it wants the supplier to produce. These potential returns also co-determine what the supplier can expect to get out of the relation, along with its interpretation of the requesting buyer's loyalty. Finally, the agents are given the opportunity to switch to an alternative partner, if it comes along. This decision is also made on the basis of the potential returns in the new relation and the probability that they will actually be realized.
In general, buyers and suppliers make decisions in managing their relations on the basis of the economic criterion of potential profit as determined by specialization of products demanded and efficiency of production and the (social-)psychological criterion of the partner's loyalty. The paper reports simulation studies that compare different strategies for managing relations, where strategies are characterized in terms of the relative emphasis on each of the two criteria.
REFERENCES:
[Albin & Foley 1992] Albin, P. & D.K. Foley (1992) Decentralized, dispersed exchange without an auctioneer: a simulation study. Journal of economic behavior and organization, 18 (1): 27--51.
[Epstein & Axtell 1996] Epstein, J.M. & R. Axtell (1996) Growing artificial societies: social science from the bottom up. Washington, DC: Brookings Institution Press and Cambridge, MA: The MIT Press.
[Porter 1985] Porter, M.E. (1985) Competitive advantage: creating and sustaining superior performance. New York: The Free Press.
[Tesfatsion 1997] Tesfatsion, L. (1997) A trade network game with endogenous partner selection. In: Amman, H. et al. (eds.) Computational approaches to economic problems. Dordrecht: Kluwer Academic Publishers. 249--269.
[Vriend 1995] Vriend, N.J. (1995) Self-organization of markets: an example of a computational approach. Computational economics, 8 (3): 205--231.
[Williamson 1985] Williamson, O.E. (1985) The economic institutions of capitalism: firms, markets, relational contracting. New York: The Free Press.
DISTRIBUTION OF COMPETITION IN ORGANIZATIONAL MARKETS.
Jeroen Bruggeman and Ivar Vermeulen
CCSOM
Organizational Ecology (OE) is a theory about ``Darwinian selection'' in populations of organizations (Carroll and Hannan, Organizations in Industry, 1995). A population, or market, consists of organizations of similar forms. Examples of such forms are automobile manufacturers.
OE attempts to explain empirically observed processes in populations, by relating population level variables. The theory fragment ``density dependence'' explains founding and disbanding rates of a population as consequences of competition and legitimation. A less well developed theory fragment, ``resource partitioning,'' relates resource partitioning proper (differentiation of niches), market concentration, and the outflow of middle sized organizations, without showing clearcut causal relations.
The two theory fragments are not related, although, in all likelihood, the variables are. In order to gain new insights into populations of organizations, we relate the variables of two theory fragments. A problem is, though, that the current population level measure of competition, used in density dependence, does not make it possible to describe intra-populational differences in competitive pressure, necessary to explain the processes in resource partitioning.
We therefore start looking at competition at the micro level, as a relation between two individual organizations. In our definition of competition, the competitive pressure that organization X exerts on organization Y depends on the substitutability of the products or services of organization Y by those of X, and on the relative sizes of X and Y (our paper gives a more extensive explanation and examples). Notice that the competition between two organizations can in our setup be asymmetric. The strength of these competitive ties can be measured in principle, and all ties in the population can subsequently be aggregated to a measure for competition at the population level.
At the population level, our definition of competition, implemented in density dependence, yields the same predictions about the founding and disbanding rates. Furthermore, once we add an empirical generalization about the size-distribution of organizations in a population, plus some assumptions from resource partitioning, we obtain new insights in the processes that underlie resource partitioning. We can show that at some point, new (and usually small) entrants in the population are forced to produce exclusive products for the market, which increases the mortality rate of the medium sized organizations, while the largest organizations remain unaffected. As a result of this, concentration increases, which, in its turn, increases the necessity for new organizations to make exclusive products. Resource partitioning proper and concentration mutually reinforce each other, until a stable equilibrium has been reached. Last but not least, we show that presumed key concepts in resource partition, like ``center of the market,'' ``niche,'' and ``generalism,'' do not have any additional explanatory value, and can be omitted for the sake of parsimony.
Our mathematical formalizion of resource partitioning, built on top of the already formalized density dependence fragment, makes it possible to proof or calculate our claims, plus the theory's current predictions.
THE ROAD NOT TAKEN: SIMULATING PATH DEPENDENCE IN ORGANIZATIONAL EVOLUTION
Anjali Sastry
University of Michigan Business School
Why do people and organizations choose one strategy or course of action over othersand then resist pressures to deviate from or revise the original choice? How do apparently small, random events shape organizational lives? How do the processes at work in organizational founding, change, learning, uncertainty reduction, and innovation contribute to self-reinforcement? In this paper, I address these questions by exploring a multi-level theory of path dependence in the evolution of organizational systems. Because the theory is complex, involving the interaction of beliefs, knowledge, structure, and strategy, I use a simulation model to explore its behavior over time. The integrated path dependence theory explains how organizations embark on and remain committed to actions that become self-reinforcing. In addition, the present research suggests ways in which organizations can break out of dysfunctional behavior patterns in which they seem stuck.
My aim in this paper is to explore systematically an integrated theory of organizational path dependence. To build my own integrated model of path dependence in organizations, in an earlier paper a colleague and I critiqued, extended, and synthesized a variety of traditionally disparate theories, all of which explain why organizationsor collections of organizationsconverge on approaches which then become difficult to change. For instance, the research on learning by doing, imprinting, and escalation of commitment offers a variety of processes by which early choices or early events in an organization are reinforced over time. The theory of normative isomorphism provides one mechanism whereby organizations within a field begin to resemble each other over time. Mimetic isomorphism, bandwagons, fads, and diffusion of innovation result from similar processes of learning from others, with the result that the organization more closely resembles others in its field. Finally, we looked to research on coevolution for ideas about how change in the organization is constrained and shaped by the interaction of multiple systemsthe social, cognitive, and material. In constructing an organizationally-grounded theory of path dependence, we specified a variety of path-dependence mechanisms that complement the increasing returns to scale and network externalities that economists have identified with path dependence. By mapping out the processes at work in each of these eight phenomena and juxtaposing the causal maps, we showed that several processes are common to many theories; we also highlighted the multi-level nature of path-dependent processes in organizations. In addition, we found that, while many theories are designed to explain how path dependence is initiated, few explain how path dependence ends.
I designed the present study to extend this earlier work and to address questions we uncovered in the process of mapping and comparing the existing explanations of path dependence. Thus, I am exploring and building on the integrated theory of organizational path dependence. The formal model that is our focus here includes over a dozen separate self-reinforcing processes. By formalizing them, I can explore their interaction and evolution over time. Among the key questions is one about synthesizing levels of analysis. While we see that such theories as escalation of commitment contain multiple levels of analysis, the relationships between the levels are unclear. How do phenomena that operate at the field level emerge from organizational-level processes? Can psychological explanations at the individual and interpersonal level explain organizational path dependence? The simulation model allows us to explore these questions, as we can alter the effects of each of the numerous self-reinforcing processes to establish which elements are necessary to generate path dependence. Using the model, I also explore how the timing of noise and variation in an organizations decision making affects its evolution, as random events have been argued to exert a larger effect on young organizations than on older ones. I also investigate how the processes active in path-dependence interact with other processes in organizations, such as goal-seeking. Finally, I test a range of ideas for breaking out of path dependence traps.
ORGANIZATIONAL ADAPTATION
Kathleen M. Carley
Carnegie Mellon University
Social and Decision Sciences & H.J.Heinz III School of Policy and Management
and
JuSung Lee
Carnegie Mellon University
Social and Decision Sciences
Organizations can and do adapt their designs over time. Many of the ways in which organizations change are not adaptations, i.e., they do not serve to maintain or improve performance. The question is: Which changes represent adaptations? Or in other words: How should the organization adapt to environmental and task changes? Organizations often find the need to alter their design in response to changes in the environment. Changes in the environment include such things as certain resources become unavailable or are used up, the specific task being performed changes, the rate at which new problems arise changes, the types of problems faced change, and so on. In response to such situations the organization may alter its design. Indeed, it is often felt that if organizations do not dynamically respond to a changing environment then organization performance will suffer as the design becomes outmoded. Such structural changes may be pre-planned as when switching from one well defined architecture to another, or they may be evolutionary as in automatically adapting in response to some local stimuli. Such changes may adding or dropping personnel, reassigning who is doing what task or has access to what resources, or reassigning who is reporting to whom. However, not all such changes are adaptations; i.e., not all such changes serve to improve performance or even maintain current levels of performance. Some changes, e.g., might interfere with the ability of personnel to learn and to make use of their previous experience, thus degrading the performance of the entire organization. There may be little time for the CEO to ponder the nuances of the situation and locate the absolutely optimal design. Rather, the organization must respond rapidly, make do with what it has, and satisfice. There are typically, in place and often unalterable, constraints (particularly in the short run) that the organization must work around. In other words, the organization is faced with a problem in constrained adaptation. There is little understanding of what types of change represent adaptations, particularly when the organization must operate within constraints.
Literature on organizational adaptation suggests that organizations do change over time [Stinchcombe, 1965; DiMaggio and Powell, 1983; Romanelli, 1991]. Part of this change is due to strategic re-organization [Kilman and Covin, 1988] including re-engineering and re-organization. However, not all types of re-organization may be equally valuable. For example, organizational performance may improve as individual members of the organization gain experience [March, 1981]. But, replanning which moves personnel between positions and reassigns them to different tasks may result in loosing the benefits of the experience the members had gained previously. Little is known about organizational change, and even less about how organizations should be designed to promote adaptation. Most theories of organizational design speak to the relative advantage of different designs in different situations [Lawrence and Lorsch, 1967; Hannan and Freeman, 1977]. Such theories, in principle, provide some guidance for organizational change. For example, population ecology can be interpreted as suggesting that if the organization is moving out of a niche environment then the organization should move from a specialist to a more generalist structure [Hannan and Freeman, 1977]. As another example, Staw, Sanderlands and Dutton [1981] have argued that organizations when faced with a decrease in their performance will shift to a more rigid and centralized structure such as is typical in many hierarchical forms. Such suggestions, however, provide little theoretical, let alone practical, guidance as to how to set up the organizations structure so as to facilitate adaptation. The ultimate goal should be to develop some practical guidelines for adaptation.
Theorizing about organizational adaptation is difficult. The dynamics of change result from simple, but possibly non-linear processes. Consequently, thinking through the implications of adaptation processes is non-trivial. Consider that the following two illustrative processes, retasking and individual learning, may occur simultaneously. When performance drops organizations may enter a downward spiral by choosing to move personnel to work on tasks using resources with which they have little experience, thereby loosing the benefits of the experience these same personnel had in other tasks and with other resources, which in turn may lead to a further reduction in performance, which may lead to further retasking. Alternatively, such retasking may make it possible to perform more tasks simultaneously thereby increasing the speed of organizational response and possibly increasing overall performance regardless of individual expertise or training. Given just these two processes, retasking and learning, what will be the impact of change? Will the retasking be an adaptation? How can issues such as these be addressed?
In this paper, the question of organizational adaptation in a changing environment is addressed using computational analysis. Using a computational model of organizational performance, a series of virtual experiments were run to examine which changes in design were adaptations. A series of virtual experiments were run using a computational model of organizational performance, ORGAHEAD [Carley, 1996; Carley & Svoboda 1996]. In this model, organizations have the ability to change aspects of their design and the nature of such changes and their relation to performance is monitored. This model couples a standard model of individual experiential learning with a model of organizations as strategic adaptive agents. At the individual level learning is carried out using a standard experiential learning model based on work in cognitive psychology. At the organizational level adaptation is carried out using a simulated annealing model. The simulated organizations were given a basic set of tasks to perform, then this set of tasks was dynamically switched to another set one or more times. Results indicate that frequent change can be maladaptive. An organization is said to be adaptive if the changes that it makes to its design serves to maintain or improve its performance.
ORGAHEAD is a multi-level information processing model of strategic and experientialadaptation in which the organization can change its structure and the agents can garner experience. At the operational level organizations are characterized as being composed of a set of agentsarranged in some type of command structure. The organization is facing an environment that may change in various ways. The agents are adaptive agents, each of whom occupies a particular position and has the capability of learning over time as experience is gained with the task being performed and the resources they are using. Organization level performance is determined by the actions of the individuals in the organization as they work on tasks. The specific model used is the CORP model of organization performance. At the strategic level, the organization can adapt strategically in response to changes in its performance by altering its design in a number of different ways including personnel movement, retasking, and reassignment of personnel. Organization performance is affected by the ability of the CEO to anticipate the future and take the appropriate strategic actions to alter the structure in response to actual or anticipated environmental cues. This strategic adaptation is modeled as a simulated annealing process.
ORGAHEAD has been informed by empirical studies both on individual learning by humans and on adaptation within human organizations. ORGAHEAD is a very versatile program in which the user can specify the initial structure of the organization (or set of organizations), whether agents employ SOPs or experience in making decisions, how much training agents receive, how much agents remember, the type of strategic changes allowed, the initial likelihood of the allowable changes, the maximum frequency of change, the rate at which the organization becomes risk averse, the "function" the organization is trying to optimize (e.g., performance or performance subject to minimizing communication), the task environment, and several types of "triggers" for change (such as change in task environment or destruction of resources).
In ORGAHEAD simulated annealing is used to capture the strategic constraint based adaptation process that the organization goes through. Over time, the organization (more precisely the CEO) attempts to optimize the design given some cost function. The cost function depends on the organization's goal; illustrative goals include maximizing decision accuracy, maximizing kill ratio, minimizing communication. The CEO alters the design strategically; that is, a change is made if it appears to move the organization closer to the goal regardless of whether or not it actually does so [Simon, 1944; March and Simon, 1958]. The organization (more precisely the CEO or central organization) is not omniscient, does not compare all strategies, but simply evaluates a strategy through a kind of "what if" analysis, trying to forecast or anticipate, albeit imperfectly, the future [Allison 1971; Cohen and March 1974; Axelrod 1976]. Since the forecast is known to be imperfect, the CEO may at times gamble on redesigns that might possibly "increase costs" if it is felt that there is some long term advantage. Overtime, the number of high risk moves decreases [Stinchcombe, 1965b] as the organization locks into a certain way of doing business and so gets trapped by its competency [Levitt and March, 1988].
A series of virtual experiments were conducted using ORGAHEAD. These experiments were designed to examine the ability of the organization to adapt given different types of environmental changes. The level of volatility and nicheness of the environment were varied. Under each experimental condition 1000 organizations are simulated, using the Monte Carlo technique. Each organization starts with a design. Each organization's initial design was chosen at random. Each organization is simulated for 20,000 tasks. Changes in the design are allowed to occur every 500 tasks. When the CEO thinks about a proposed change, the "hypothetical organization" is "simulated" for 250 tasks and its performance measured on these 250 tasks. This expected performance is then contrasted with the latest actual performance. Actual organization performance is measured every 500 tasks. When the organization changes its design, it is done so on the basis of the expectation that the new structure will be a better performer, rather than on actual feedback from previous exercises. Thus, some of the changes that are taken may actually, particularly in a changing environment, degrade performance. Each agent in the organization has a maximum memory capacity of 250 tasks; the agent's behavior reflects what it had learned from solving the last 250 tasks.
Results indicate that frequent change does not guarantee adaptation. Indeed, whether we look at stable environments or any other type of environmental conditions, the adaptive organizations (top performers) make fewer changes in personnel than do the non-adaptive organizations.In stable environments, adaptive organization increase in size whereas non-adaptive organizations decrease in size. While both adaptive and non-adaptive organizations decrease in density the adaptive organizations show a greater level of decrease. In other words, adaptation is increasing size and decreasing interconnection by creating a system that decentralizes the organization in terms of resolution of low level conflicts; i.e., least upper boundedness decreases.These same patterns are seen regardless of the environmental conditions. That is, whether the change is coarse-grained or fine-grained or a single step adaptation does not equal frequent change. Interestingly, adaptive organizations in a fine-grained environment do exhibit more frequent change than do those in a coarse grained environment; though, not as many as in a step environment. Also in a stable environment the number of changes is lower still. Further, adaptive organizations tend to increase in size and dramatically decrease their density.
REFERENCES
[Allison, 1971] Graham, Allison, 1971. Essence of Decision. Boston, MA: Little Brown.
[Axelrod, 1976] Robert M. Axelrod, Structure of Decision: The Cognitive Maps of Political Elites. Princeton, NJ: Princeton University Press, 1976.
[Carley & Svoboda, 1996] Kathleen M. Carley & David M. Svoboda, "Modeling Organizational Adaptation as a Simulated Annealing Process." Sociological Methods and Research, 25(1): 138-168, 1996.
[Carley, 1996] Kathleen M. Carley, "Adaptive Organizations: A Comparison of Strategies for Achieving Optimal Performance" in Proceedings of the 1996 International Symposium on Command and Control Research and Technology. June. Monterray, CA, 1996.
[Cohen and March, 1974] Michael D. Cohen, and James G. March. Leadership and Ambiguity: The American College President. New York: McGraw-Hill, 1974.
[DiMaggio & Powell, 1983] Paul J. DiMaggio and Walter W. Powell. "The iron cage revisited: institutional isomorphism and collective rationality in organizational fields." American Sociological Review, 48: 147-160, 1983.
[Eccles and Crane,1988] Robert G. Eccles and Dwight B. Crane. Doing Deals: Investment Banks at Work. Boston, MA: Harvard Business School Press, 1988.
[Hannan & Freeman, 1977] Michael T. Hannan and John Freeman. "The Population Ecology of Organizations." American Journal of Sociology 82: 929-64, 1977.
[Kilman and Covin, 1988] R.H. Kilmann and T.J. Covin (Ed). Corporate transformation: revitalizing organizations for a competitive world. Vol 1. in The Jossey-Bass management series, San Francisco, CA: Jossey-Bass, 1988.
[Lawrence & Lorsch, 1967] Paul R. Lawrence, and Jay W. Lorsch. Organization and Environment: Managing Differentiation and Integration. Boston: Graduate School of Business Administration, Harvard University, 1967.
[Levitt & March, 1988] B. Levitt, and J. March, J. "Organizational Learning." Annual Review of Sociology, 14: 319-340, 1988.
[March and Simon, 1958] James G. March, and Herbert Simon. Organizations. New York: John Wiley & Sons, Inc., 1958.
[March, 1981] James G. March, "Footnotes to Organizational Change." Administrative Science Quarterly 26: 563-577, 1981.
[Romanelli, 1991] E. Romanelli, "The Evolution of New Organizational Forms." Annual Review of Sociology 17: 79-103, 1991.
[Simon, 1944] Simon, Herbert A. "Decision-making and Administrative Organization." Public Administration Review 4: 16-31, 1944.
[Staw, Sanderlands & Dutton, 1981] B.M. Staw, L.E. Sanderlands and J.E. Dutton. "Threat-Rigidity Effects in Organizational Behavior: A Multilevel Analysis." Administrative Science Quarterly , 26: 501-524, 1981.
[Stinchcombe, 1965a] A. Stinchcombe, "Organization-creating organizations." Trans-actions, 2: 34-35, 1965.
KAUFFMAN'S NK MODEL OF COMPLEXITY AS CAUSE: A CRITIQUE
Bill McKelvey
UCLA Anderson School
**** See End ****
REAL-TIME HIERARCHICAL RESOURCE ALLOCATION
Timothy Van Zandt
Department of Economics
Princeton University
Consider the capital budgeting processes of large firms, or the procedures for allocating resources within large non-market organizations such as governments, firms and universities. The flow of information in these procedures may be hierarchical and resemble the following scenario. At the bottom of the hierarchy are the operatives or shops or whatever are the ultimate recipients of resources. In the upper tiers are managers or administrators, who are independent of the shops. Information about the operatives' costs or needs is aggregated by a flow of information up the hierarchy, and resources are recursively disaggregated by a flow of information down the same hierarchy. These procedures exhibit both decentralized information processing, meaning that the resource allocations are calculated jointly by the members of the administrative staff, and decentralized decision making, meaning that each node makes decisions which constrain the resource allocations and that the decisions of different nodes of the hierarchy are calculated using different information.
The purpose of this paper is to construct a model that distinguishes between these two forms of decentralization and that explains the advantages of each of them, particularly the decentralization of decision making. We achieve this by modeling the allocation of resources as a learning problem in which the decision rules are calculated in real time, and by modeling the boundedly rational agents who calculate the decision rules as random access machines, as in Radner and Van Zandt (1992) and Van Zandt and Radner (1996). We are then able to represent decision procedures that have the hierarchical upward and downward flows of information described above. The nodes of the hierarchies correspond to multi-person decision-making units (offices) within which there is decentralized information processing and aggregation of cost information. It is the disaggregation of resource allocations (the decentralized decision making) that defines the hierarchical structure.
In the model, the advantage of decentralized information processing, whether it is within or across nodes, is that operations can be performed concurrently by several agents and hence delay is lower than when one person performs all the operations sequentially. The reduction in delay, in turn, means that resources are allocated based on more recent information.
When resource allocations are disaggregated through the hierarchy, each office suballocates resources to its subordinates based only on the aggregate of these subordinates' cost information. Because each office's information is less aggregated than the aggregate cost information used by its superior, it is also more recent. Thus, the value of the decentralized decision making is that managerial nodes in the lower tiers can allocate resources within small groups using recent information, while nodes in higher tiers are still able to exploit gains from trade between the groups, although based on older information. The main cost of this decentralization is an increase in the amount of computation that is done.
The advantage of decentralized information processing mentioned above also appears in models of decentralized batch processing. Batch (off-line) processing, even when there are multiple arrivals of unrelated processing tasks over time, is different from the real-time (on-line) learning in our model because in the former all information for a computation task is available at the same time and there is a single moment when the computation is completed, whereas in the latter, decisions are made repeatedly and there is a flow of new data that are potentially relevant to decisions made at multiple epochs. Hence, with real-time learning decisions are typically based on data of heterogeneous lags and there is no single measure of delay. This distinction is important for the decentralization of decision making in our model, which is valuable because of the heterogeneity of the age of information across nodes. We show that, in contrast, a decentralized batch processing model cannot explain the decentralized decision making described above; in such a model, the hierarchical disaggregation of resource allocations actually increases both delay and computation costs.
In one version of the model that we study, each operative has a simple quadratic cost function whose cost parameter follows a stationary AR(1) process that is independent of the parameters of the other operatives. The result is a surprisingly tractable model that we can use to characterize optimal procedures within a restricted class, to study returns to scale, and to perform comparative studies with respect to the speed at which the environment changes.
This quadratic model also allows us to see how decision rules, though constrained by the computation technology, can take into account the statistical inference problem posed by the changing environment. We use the static team theory model of hierarchical resource allocation in Geanakoplos and Milgrom (1991) as a tool for suggesting decision rules that take into account the statistical assumptions and for deriving the expected shop costs for these rules.
We obtain the following results on returns to scale. If the number of tiers is fixed, then the average cost curve is approximately U-shaped even when the managerial wage is zero, because of the decision-theoretic cost of computational delay. However, when the number of tiers in the optimal hierarchies is allowed to increase with the number of operatives, computational delay alone does not inexorably lead to eventually decreasing returns to scale. This is because expected shop costs can always be reduced by joining independent hierarchies under a central office that coordinates allocations to these hierarchies. Nevertheless, because of cumulative delay, the value of this coordination exceeds the administrative cost when the wage is positive and the hierarchies are large. Therefore, managerial wages combined with delay lead to a bounded firm size. Furthermore, we find that firms are smaller and more internally decentralized the more rapidly the environment changes.
DESIGNING QUALITY INTO PRODUCT DEVELOPMENT ORGANIZATIONS THROUGH COMPUTATIONAL ORGANIZATIONAL MODELING AND SIMULATION
Jan Thomsen, John C. Kunz, Yul K. Kwon, Sam Miller,& Raymond E. Levitt
Stanford University
This presentation discusses our link of Total Quality Management (TQM) theory with theory and practice of computational organizational modeling and simulation. We applied organizational behavioral theory to TQM and built a unified theory of organization quality. Implemented in a computer simulation, the model measures and predicts useful and measurable aspects of organization quality. We prospectively applied the simulation model early in the design of an industrial test case and made recommendations based on our analysis. Considering our recommendations, the cooperating industrial manager intervened in the engineering process to reduce some of the organizational risks that we predicted might adversely impact project performance. In our subsequent observations of the project, the potential risks did not appear. This work contributes to computational organizational modeling, TQM and organizational modeling methodology.
Background
The increasingly competitive environment that organizations must face, particularly as a consequence of the rise in global competition, has necessitated a progressively stronger emphasis on quality management within organizations. Prior to the 1960=92s, assessments of quality primarily considered measures of product quality as indicated by such variables as dimensional tolerances and the number of functional defects. Management generally did not attempt to evaluate quality through internal mechanisms, but relied on customer response and feedback. Unfortunately, this passive approach to quality control was costly and inefficient, since the activities and work processes that gave rise to products were already well established by engineers once customers identified quality problems.
After the 1960's, the focus of quality control shifted from measuring the outputs of production to monitoring the work process. Early efforts in this approach to quality management used Statistical Process Control to measure and control variance within the production process.
Although organizations now possess methods and criteria to evaluate work process quality, managers still lack methods to anticipate how changes to the organization or to the work process will affect production quality, cost and speed. Beyond relying on their own experience and intuitions, decision-makers cannot systematically predict how different organizational structures, communication tools, personnel qualifications or work processes, will promote or degrade quality. The challenge facing organizational decision makers and organizational researchers today is to design quality into organizations instead of developing further ways to improve quality after deficiencies have already affected the product. Their task is complicated by the fact that increases in quality often require a trade-off in other performance measures, such as production costs and durations.
Our research shifts the focus of quality management one step further upstream from the process toward measuring and controlling the quality of design and development organizations.
Our computational model operationalizes an information processing view of the product development process. Product development work consists of a process in which designers search a space of potential solutions to find a solution that best realizes a number of desired goals concurrently. For this research, we consider exceptions to be deviations from managerial prescriptions. We define three types of exceptions: Technical errors are mistakes made by an actor that are of a technical nature. They are categorically detrimental to the product quality or successful accomplishment of the desired objective and must be corrected in order to ensure the reliability and functionality of the product being developed.
Productive non-conformances are beneficial non-conformances that represent solutions that are superior to the ones anticipated by the manager or the project plan i.e., they achieve a more desirable trade-off among project goals. Counterproductive non-conformances represent alternative accepta ble solutions that are inferior to those anticipated by the manager.
The probability that an actor generates a technical error depends on the competence of the actor vs. the complexity of the task being executed. The probability that an actor generates a non-conformance depends on the relationship between the actor and the supervisor as well as the potential size of the solution space of the associated activity. The probability that a non-conformance is productive or counterproductive depends on the relative skills of the manager and the subordinate. For example, a relatively unskilled manager will encounter more productive non-conformances from a highly skilled subordinate than one who has relatively low skills.
In our model, the occurrence of these three types of exceptions measure organizational quality. These quality measures, combined with additional measures for project cost and duration, provide metrics for evaluating the effectiveness of different organizational designs.
Design proceeds in synthesis-analysis-resolution stages (Coyne et al., 1990). The organizational quality objectives change with these stages.
Group heterogeneity facilitates synthesis of ideas and solutions as different perspectives, skills and backgrounds of team members will normally promote diversity of ideas. We measure of the quality of the organization that performs synthesis by the number of new productive ideas that the team generates.
In the analysis stage, team members attempt to achieve a common understanding concerning which options are likely candidates for implementation. We measure quality of the organization that performs analysis as the number of completed communications compared with total number of communications.
The final stage of development is resolution, in which a decision is made and a solution is selected. We assess the quality of the organization that performs resolution by the number of decisions that are made properly and on time through the appropriate decision-making process.
The relative importance of organizational quality for each of these phases will vary by industry. Firms in industries that place a strong emphasis on new product development, such as integrated microchip manufacturers, will generally have a greater need for quality in the synthesis stage rather than in analysis or resolution. In contrast, industries that produce mature products will give more weight to analysis and resolution than synthesis. For example, it is more important to a manufacturer of toilet paper to ensure reliability rather than to promote product innovation.
The detailed level of granularity at which our model simulates organizational behavior allows us to measure quality for each stage of development. In addition, our model can predict where in the development process quality problems may appear. For example, our simulation produces a visual representation of time-varying activity execution that allows a user to identify bottlenecks in the work process. Another graph shows the in-tray depth (i.e., the time-varying number of items that await an actor's attention) and can be used to predict when and where certain actors become overloaded.
Our computational model uses and extends the framework developed by the Virtual Design Team (VDT) (Jin and Levitt, 1996) research group at Stanford University. We applied our model to a product development team within an aerospace firm charged with the design of a new specialized valve used in satellites.
This project included a prospective intervention study. We presented preliminary results of our model to a cooperating industry project manager. The results predicted potential future bottlenecks in the work process and suggested that organizational quality would be significantly affected by changes in organizational parameters and in the characteristics of personnel. The project manager for the product development team considered our results, discussed possible corrective actions, and decided to invest in greater monitoring efforts in order to reduce the likelihood for future quality problems in the project. The project subsequently performed exceptionally well. We cannot attribute project success to the manager's intervention, however we are encouraged by both the manager's having decided to intervene following analysis of our model and by the subsequent favorable results.
MARKET ORGANIZATIONS FOR CONTROLLING SMART MATTER
Oliver Guenther, Tad Hogg and Bernardo A. Huberman
Xerox Parc
We have developed a distributed control system for smart matter that is based on an underlying market organization among sensors and actuators trying to stabilize an inherently unstable physical structure. This is achieved by focussing control forces in those parts of the system where they are most needed. This is because computational markets can successfully coordinate asynchronous operations in the face of imperfect knowledge and a changing environment. Moreover, as in economics, the use of prices provides a flexible mechanism for allocating resources with low information requirements.
In the market control treated here, actuators are treated as consumers and the external power sources are the producers. Each consumer communicates with several other consumers. The range of communication is given by the organizational structure of the market. We focused on four different structures and examine how a simple learning algorithm that allows the system to change its structure can improve the system's performance.
Our simulations show that the control performance depends on the used market structure and that a hierarchical and a multihierarchical structure are a reasonable compromise between rapid local responses with simple communication and the use of global knowledge which can often not be assumed. Moreover, letting the consumers explore whom to communicate with leads to novel organizations that use less power than the ones we studied above.
COMPUTATIONAL SUPPORT FOR LOGICAL FORMALIZATION
Jaap Kamps
Applied Logic Laboratory,
Center for Computer Science in Organization and Management,
University of Amsterdam,
Social scientists usually agree that theories should be logical, but they rarely address the issue, eschewing the difficulties of investigating the logical structure of `discursive theories' (theories expressed in natural language, the standard representation in the social sciences). It is part of the research program at the Applied Logic Laboratory (ALL) to tackle this problem head on by `formalizing' theories from the social sciences, that is, by rationally reconstructing them and expressing them in logical form, e.g., P'eli et al. (1994). In this endeavor, computational support is invaluable. At ALL we use automated theorem provers (ATPs), such as the well-known Otter (McCune 1994b), as well as automated model generators (AMGs), such as Mace (McCune 1994a). These tools can test for logical criteria that correspond to natural questions which we would like to ask about the theory:
- Is the theory contradiction-free? (in logical terms, is the theory logically consistent?) If an AMG can construct a model of the premise set then the theory is logically consistent. If an ATP can derive a contradiction from some subset of the premises, the theory is inconsistent.
- Is the argumentation of the theory fallacious? (is a `theorem' in fact a false conjecture?) If an AMG constructs a model where the premises hold but the conjecture does not hold, the conjecture is refuted. Moreover, examining such a counterexample is usually an indispensable step towards a `real' theorem. Proving soundness can be directly checked using an ATP.
- What unstated background knowledge is necessary? In any exposition of a theory, a certain amount of background must be `taken for granted,' that is to say, it is assumed to be indisputable as common knowledge. But such information must be explicitly added to a formalization. Again, an AMG can be used to find a counterexample, which will immediate reveal the missing knowledge.
- What assumptions has the author neglected to make? In contrast to the common knowledge of the previous point, occasionally the formalization of a theory reveals a genuine hiatus in the theory. In this case, counterexamples generated by an AMG can not be as easily discarded: we have to deal with a genuine counterexample. This may necessitate either strengthening the premises, weakening the `theorem', or even discarding the `theorem' altogether. Examining the counterexample(s) provides useful information for deciding between options to refine the theory.
- What is the domain that the theory describes? (what do the models of the theory look like?) An AMG can be used to provide candidate models for exploration. This gives insight into the domain which the theory describes.
- Is the theory falsifiable? (are the theorems logically contingent?) If no state of affairs can possibly falsify a theory, then it is a waste of time to empirically test the theory. For this reason, falsifiability is an essential property of a scientific theory. If an AMG can construct a model (disregarding all premises) where a theorem of the theory is false, then this theorem is falsifiable. If an AMG can also construct a model in which the theorem holds, then this theorem is satisfiable too. A theorem that is both satisfiable and falsifiable, is contingent: the validity of the theorem is strictly determined by the premises.
Theory formalization in the social sciences has been revived by the powerful methodology of logical formalization (Hannan 1997). In this approach, computational support is essential to vivify the logical formulas: formal logic provides criteria for scientific theories that can be evaluated using automated reasoning tools, such as ATPs and AMGs. The case studies at ALL/CCSOM show how the logical criteria and computational tools can play an important role in the formal (re)construction of scientific theories. On the one hand, these criteria provide logical tests for scientific theories. On the other hand, failing a test will provide useful information for refining the theory.
REFERENCES:
Hannan, M. T. (1997). On logical formalization of theories from organizational ecology. In A. E. Raftery (Ed.), Sociological Methodology 1997. Blackwell, Oxford UK.
McCune, W. (1994a). A Davis-Putnam program and its application to finite first-order model search: Quasigroup existence problems. DRAFT.
McCune, W. W. (1994b). Otter: Reference manual and guide. Technical Report ANL-94/6, Argonne National Laboratory, Argonne, Illinois.
Peli, G., J. Bruggeman, M. Masuch, and B. 'O Nuall'ain (1994). A logical approach to formalizing organizational ecology. American Sociological Review 59 (4), 571-593.
P-CALCULUS AND THE LOGIC OF ORGANIZATIONS
Corrado Pasquali
University of Genova
In my presentation I wish to illustrate p-calculus and propose it as a computational model for the notion of organization and for (some) organizational phenomena. p-calculus is an abstract model of interaction and concurrent computation that overcomes some of the mathematical and conceptual limits of other models of concurrency such as Petri Nets, CCS, CCP and ACTORS. My main thesis is that, first: computational models of organizations should be based on concurrent computation as this can be used as a general, abstract theory of organization; second: p-calculus can be interpreted as a calculus of arbitrary configurations of processes and can be easily thought as a modeling tool for organizational routines and other organizational phenomena.
An organization (and, broadly speaking, any system) has two fundamental features: diversity (it is made of different, recognizable subcomponents) and unity (it is a recognizable, global entity). This latter feature is grounded on the way the components of an organization are permitted to interact with each other. In this sense, an organization is nothing but a set of constraints on interactions: a configuration of units interacting under specified (and possibly changing) constraints. It is defined by individuating a class of interactions out of a multitude of possible ones thus producing a global behavior from local interactions. In other words, an organization is the class of emergent properties of a set of particularly constrained interactions.
Not only an organization is a recognizable global unity made by the set of constraints on interaction that defines it, but it is also an entity that is capable of evolution, adaptation, self maintenance and that has its own dynamics. Now, the main point is that these features are shaped by the same constraints on interactions that structurally define the organization. Thus it appears to be a precise structure/action/evolution relation that bounds the dynamic flows of an organization in its evolutionary space. This is most evident, for instance, in the case of economic organizations such as firms whose organization can result in different learning possibilities or different capacities of adapting and reacting to changing market conditions. As we shall see, this problem can be dealt with in the more abstract case of information processing organizations where to different ways of organizing information processing resources, of constraining information flows and of generating different configurations of processing resources, different success, efficiency and properties of problem solving may correspond.
Concurrent programming is probably the most interesting step towards abstraction not only in programming languages but, broadly speaking, in information processing models. In concurrent programming an abstraction is made on the execution order of programs. While in Von Neumann architectures perfect sequentiality is required and a total ordering is defined on the class of processes that a program is made of, in concurrent programming this no longer holds: only a partial order is defined on processes so that these are allowed to overlap, be executed in parallel and interact with one another. So while in sequential programming there is no interaction between processes1 (each process must terminate before the next one can start), in concurrent programming this is not always the case and processes can interact with each other in various ways establishing relations of interdependence, competition, cooperation. The main problem is thus constraining interactions and execution in such a way that only a subclass of the possible execution orders is inducted in order to satisfy the desired properties of a program. Once again, this brings the notion of configuration back into the realm of analysis: we are faced with the problem of defining a configuration of processes that can solve a given problem (i.e. that satisfies the properties of a program). In my opinion, this is the sense in which concurrency theory can be thought as the general theory of abstract organizations. The core problem in concurrency is indeed nothing else than an instance of the organizational problem of constraining information flows and interaction so to satisfy certain properties and induce a preferred dynamic path. 2 However, what really matters is that concurrency theory offers the possibility of a precise mathematical and computational analysis of the notion of organization and of (at least certain classes of) organizational phenomena in purely abstract and process-theoretic terms.
The more common toy example used to give a hint of what concurrency is, is that of two programs P1 and P2 that are so defined:
P1: x : = 1; x : = x + 1
P2: x : = 2
When executed, these two programs both transform the initial memory by replacing the value of x by 2 and thus they have the same meaning. But suppose now that you have a t hird program:
Q: x : = 3
When Q is concurrently composed (executed) with P1 and P2 the result will be:
R1: P1 =Q R2: P2 =Q3
where R1 and R2 will have different meanings depending on how the component processes will interact with each other. So, while in the sequential case there will be no quest for constraints on interaction and a compositional semantic based on the Scott-Strachey approach will do very well, in the concurrent case we need a semantic that can take into account the ways in which a program interacts with the memory. That is: we need a semantic in which interaction is treated as the fundamental category and a general framework for concurrency that can extend and go beyond the purely functional paradigm of sequentiality. This argument is similar (and very close in spirit) to one of the best known "slogan" of complexity theory and of nonlinear systems: components and their sum are not enough to explain a system: there is more to them and that "more" is interaction. Actually, while we have a whole repertoire of sophisticated and well working models for computational worlds made of functions (namely: Turing machines, l-calculus and their extensional equivalents), we do not have anything of comparable precision and sophistication for concurrency and interaction. Every model since proposed such as Petri Nets, CCS, CCP, ACP or ACTORS has its own limits. In my presentation, I would like to show these limits and the way they are solved by Milner's p-calculus; at the same time I would like to argue for p-calculus as a purely process-theoretic theory for abstract organizations.
Another point to be put forward with respect to the claim of concurrency as the abstract theory of organizations regards the meaning of computation in concurrent systems. In such systems a computation is not intended to halt and terminate at any time: their behavior is better explained in terms of conceptually infinite sequences of state transitions. Systems of this kind (such as database servers, air traffic control systems, mobile telephone network s...) continuously accept requests and stimuli from the environment and react to them changing their state and this, in turn, will affect reactions to future requests. In my opinion, it seems plausible that the focus on the "avoidance of deadlock" problem instead that on the "halting" problem puts concurrency theory very closely related to systems (such as organizations be they natural or artificial, economic or living systems) that are capable of persisting through time and preserving their identity while changing their internal structure.
In the remaining part of this abstract, I shall try to give a hint of the fundamental notions of Milner's p-calculus. p-calculus' world is only populated with processes that communicate among them by connected ports and that eventually change their structure as a consequence of communication. The notion of process supported by the calculus is that of an entity defined by an internal behavior (given by a behavior definition) plus its possibilities of interaction (given by a set of "ports" through which, links are established with other processes. A class of processes linked by ports is a configuration of processes). The main building block of the whole thing is the notion of communication and, in particular, the identification of every form of interaction with it. This reduction is based on the observation that a process can either interact with other processes (i.e. it communicates with them and changes its state consequently) or it occur independently: but even this case can be considered as an "internal communication" (e.g. a state change occurring in a computer when a variable is assigned a value by a program). So, what really matters to the end of providing a definition of "behavior" for a system of processes is a description of the ways its components can communicate with one another and of the ways the whole system can interact with the outside environment (this is very close in spirit with the definition of "organization" proposed in the first two paragraphs). Given this description and the definition of communication as the only form of interaction, it is possible to describe a system's behavior as its total capacity of communication. Milner's idea is to treat communication as a single, indivisible act: the usual categories involved in communication such as "sender", "medium", "receiver" are reduced to the single category of "performer" imagined as the participant to single, indivisible acts of communication. This permits the elimination of any passive/active entities distinctions and the substitution of "shared memory" models with a distributed and interactive model of program-to-memory interactions: in p-calculus everything is treated as a process and program variables are regarded as channels of interaction. This point has at least two important consequences; first: memory is no longer treated as a monolithic, centralized entity because (as the previous example suggests) different parts of a memory can be simultaneously accessed. Secondly, channels of interaction (represented in the calculus by ports and their complements) are not "channels" in the sense of "something that has a capacity", but "possibilities of interaction" or "contiguities" and, more importantly, they are themselves entities than can be communicated and passed through ports. Thanks to this latter feature (which is absent in other models of concurrency such as Petri Nets, CCS or CCP) in p-calculus mobility can be directly represented in processes that change their structure (i.e. their ports and possibilities of interaction) and that produce changes in the configuration of processes they are in.
In its essence, all that there is is just communication through complementary ports and processes performing state transitions and configuration transitions as a consequence of communication. The possibility of a calculus of configurations in which (constrained) interaction and the ways in which it induces (endogenous) change on a configuration of processes performing it, is, I believe, a powerful tool in analyzing a wide variety of down to earth organizational phenomena (such as social interaction and organizational routines) and a powerful though elegant and simple abstract theory of the very notion of organization.
REFERENCES:
Milner, R. et al. RA calculus of Mobile ProcessesS I & II, in: Information and Computation, 100, 1-77, (1992)
Milner, R. Communication and Concurrency, Prentice-Hall, Englewood Cliffs, NJ, 1989
NOTES
1 Actually, this is not completely correct as in sequential programming you don't have processes in a strict sense, but only functions. However I will skip a discussion on this (fundamental) theme in this pages.
2 This latter point is much more subtle and complex than it is stated here as the structure/dynamic relation is not strongly predictable and might be seen as similar (or even equivalent) to the G-type/P-type relation in
biology.
3 R=R denotes concurrent composition.
CHALLENGES FOR ORGANIZATIONAL LEARNING ISSUES IN LOGICAL MODELLING
Laszlo Polos
CCSOM
1.INTRODUCTION
Logical modelling makes things serious. Difficulties, paradoxes might as well remain hidden during both informal theoretizing and computer simulation, but appear crystal clear if we look at the phenomena from a logical point of view. In the present paper I discuss three related problems that need to be addressed by theories of organizational learning:
(1) the identity problem,
(2) the memory problem, and
(3) the non-monotonicity problem
To ensure a reasonably happy ending I also suggest (logical) solutions for all three of this problems. But while I am convinced that these problems need to be addressed, I see the solutions I suggest as -- perhaps reasonable, but --possible solutions only.
2. THE IDENTITY PROBLEM
Organizational learning is a process that changes the performance of organizations, and sometimes other properties such as market share, ownership, inertia, reliability, niche even core competencies as well. These changes challenge organizational identity, and infact the question of organizational identity proves to be a notoriously slippery business. (Think of the several attempts made to identify organizations in terms legal registration, core competencies, ownership, residual rights etc, etc.)
2.1. Logic on identity
There are several principles logic uses to characterize identity. The one that is relevant for us now is know under the name of Leibniz Law, or salva veritate principle. In short this principle says that identicals are replaceable without violation of the truth ( salva veritate). In a bit more details "a" and "b" are identical if and only if any sentence of the form P(a) remains true if we substitute "b" for "a", i.e. a=b if and only if P(a)<=>P(b).
All extensional logics, first order logic in particular, accepts the Leibniz Law, and that leads to a number of ad hoc tricks in the applications of extensional logics. (In the full version of the paper I discuss some of these tricks.)
2.2. Identity of organizations
In the light of the Leibniz Law the change of any property changes the identity of organizations, that is two organizations with different inertia are bound to be different organizations. This result is counter-intuitive and if one accepts it, the results for theories are disastrous. Organizations before and after the learning process would be different organizations.
2.3. Intensionalization
The best known solution of the Identity problem is due to G. Frege. The formal details developed to make precise his solution will be presented, here I indicate only the main guide-lines.
(a) A number of new parameters need introduced, think, for example, of (1) possible worlds
(2) situations
(3) spatio-temporal locations
We call a complex of these parameters indexes.
(b) The objects (of organization theory) are intensional entities, they exists at several indexes. (In different situations, through a period of time, etc.)
(c) Not the objects themselves but their stages, spacio- temporal slices, incarnations in different possible worlds etc. do have properties and stand in relations to one another.
(d) To objects are extensionally identical is their actual stages coincide, and intensionally identical if all their stages coincide. (The Leibniz Law, in turn characterizes only the intensional identity) This intensionalization of the objects makes it possible to identify objects even if they undergo a number of changes.
3. MEMORY PROBLEM
Intensionalization on the other hand far too liberal. In principle there is no relation between the different stages of objects (organizations). This is somewhat counter-intuitive though. The two main intuitions tell us that:
(a) There is a kind of continuity in the changes, that is neighbouring stages of organizations do not differ to much.
(b) Organizations do have memory, that is temporally ordered stages of organizations exhibit a certain process of accumulation of skills, routines etc. Organizations store a memory of their history, and it is this memory that appears in form of organizational culture.
I think that the first of these intuitions can be captured by the introduction certain topologies on both organization stages and propositions about organizations, but that goes far beyond the scope of the present paper. Here I focus on the second intuition.
3.1. Modeling organizational memory
In the paper B. Nooteboom and I presented on the last CMOT Workshop I suggested that organizational changes can be usefully modeled by a certain replacement operation, and organizations are best seen as elements of an algebraic structure called replacement system, and the organizational universe can be identified with the replacement system itself. We used this model to prove that under certain conditions tacit elements of knowledge must be observable in organizations and only saturated organizations (maximal realizations of a particular form) can be informationally transparent. Here I want to go one step further, and suggest the following model:
(a) The organizational universe remains a replacement system.
(b) Organizations become subreplacement systems, i.e. replacement systems with smaller universe closed under the same replacement operation and the components of relation.
(c) Organizations as extensional objects are incarnations, i.e. the maximal elements of the reflexive transitive closure of the components of relation in the subreplacement systems which represent the organization as an intensional object.
(d) Organizational forms are the ontologies of these subreplacement systems. This model allows us to prove two important results: All organizations have a unique form and given the set of potential components all organizational form has at least two realization, an informationally transparent (without tacit knowledge) and an informationally opaque (full of tacit knowledge). The former ones are somehow the minimal realizations of the form and the later ones are the maximal realizations.
(This theory predicts an interesting empirical claim: Streamlining an organization of a given form increases the amount of tacit knowledge in the organization!)
By modeling organizations as subreplacement systems we can show that they indeed do have memories, and the information that is stored in organizational memories is represented in form of salient replacements and substitutions. Organizations can well be extensionally identical (they have an indistinguishable appearance, incarnation), but adapt differently due to the differences in their past experiences.
In intuitive terms: the actual incarnation of an organization and the environment does not determine the adaptation process / organizational learning. It also depends on the organizational culture, and organizational culture of two organizations differ if the respective past experiences of the organizations were not identical.
4. NON-MONOTONICITY
For organizations both environmental feedbacks and the influences of cultural patterns appear in form of propositions. But the status of these two kinds of decision inputs are logically rather different. Environmental feedbacks, as much as they are known to organizations, are facts, and must be taken seriously. In inferencing, decision making, they exhibit a monotonic logic, i.e. justified decisions remain justified even if new facts become known. (For optimizing behaviour, i.e. for omniscient organizations this claim is vacuously true, anyway.)
Real life organizations have bounded rationality, and typically they do not have sufficient amount of information (and information processing capacity) to optimize. And additional source of information (for example to indicate aspiration levels, chose decision rules) is available in form of cultural patterns. Cultural patterns exhibit non-monotonic inference patterns. They best seen as default rules, rules that can well have exceptions, can also be overruled by certain more specific rules, as well as by facts (environmental feedback). In the last section of this paper I suggest a particular non-monotonic logic to tell describe how cultural patterns are used in adaptation.
5. CONCLUSION
In the paper I used the techniques of logical modeling to identify three, related problems in the formalization of theories of organizational learning, organizational adaptation. To solve the problem of identity I suggested the solution via intensionalization. To provide mathematical models for organizational memory a further specification of intensionalization was used: organizations, as intensional objects are subreplacement systems, their incarnations are the maximal elements of the reflexive transitive closure of the components of relation.
The accumulated experience was represented in form of a class of salient replacements. Organizations remain (intensionally) identical until the class of salient replacements is constant. Change in the knowledge state does change the organization but adaptation, change of the actual performance does not. The knowledge of salient replacements appeared in form of default rules and their use in adaptation in general and in decision making in particular are semantically characterized by a certain non-monotonic logic.
I also suggested, tentatively though, to tell apart optimizing and satisficing behaviour on the basis of the logic the inference patterns within decision making follow. According to this suggestion satisficing corresponds to a non-monotonic logic, and optimizing to monotonic, for example classical first order, or a type theoretical intensional logic.
LEARNING ORGANIZATIONS IN ACTION - A SYSTEM DYNAMICS MODEL OF SATISFICING SEARCH WITH ADAPTIVE ASPIRATIONS
Lars Christensen & Tore Christiansen,
Department for Strategic Research, Norway
The objective of our work is to model and analyze how organizations use learning from their past to update their expectations of the future. We address this objective by modeling the interaction between environmental influence from market and competitors, effort due to organizational action, and evaluation of performance. We formalize the model in a set of dynamic equations, which relate the effect of environment, effort and evaluation. We implement the model in a system dynamics simulation tool, and run simulations with different environmental and organizational assumptions.
The model - content and behavior
We start from a simple model of organizational action, in which effort turns intentions into performance. We introduce performance evaluation and an organizational learning cycle, which turns past performance into intentions for future action. We add a second learning cycle, in which the results of the evaluation include aspirations for future performance. This is now the well-known model of satisficing search with adaptive aspirations, and can be explained as follows -
- Performance evaluation in business involves comparison of the performance which is realized at a given point in time, with that which was required. In general, the results of such evaluation modify both organizational effort and aspirations for future performance. Examples of the former include mergers and acquisitions, TQM efforts and reorganizations. Typical example of the latter is the way in which this year's financial and operational results (performance) determine next year's budget (aspirations). The model includes both of these influences.
- By comparing required and realized performance the enterprise can decide whether or not aspirations are fulfilled. The outcome of evaluation determines the degree to which organizational effort involves search for new solutions and improved efficiency, or introduction of complacency and slack resources. If performance is above aspirations the evaluation is positive and the organization is doing better than expected. In this case the level of search is reduced and the level of slack increased. If, on the other hand, the evaluation is negative, search is increased and slack reduced.
- As long as they have the same size, organizations have a finite amount of available energy. In our model we assume fixed size, and thus that increase and decrease in slack and search is bounded. In introducing this simplification we have neglected the effects changes in size due to, for example, hiring and firing of personnel.
- In general, the performance of business enterprise is determined both by organizational effort and influence from the environment. Business is conducted in a market where market performance influences the performance of each individual enterprise. In addition, the performance of competitors influences the enterprise's aspirations for performance. If the market does well, this will in general have a positive effect on performance. If competitors do well, this tends to increase the aspirations for performance. If markets and competitors do poorly, this tends to have the opposite effect on performance and aspirations.
This makes up a model where organizational learning has been described by a set of differential equations for the variation of evaluation (aspirations and performance), effort (search and slack), and environment (market and competitors). We have modeled these equations in the Powersim? system dynamics simulation tool, using a graphical interface to define the various levels, constants and links. We use Powersim? to investigate the behavior of the model in a set of simulation runs.
The simulation - runs and results
The simulation runs are meant to represent different assumptions about environmental and organizational behavior, and the results should be interpreted as illustrations of organizational response to learning. In a closed system (without environmental input) the difference between performance and aspirations quickly disappears and all variations in organizational effort cease. Time series of system behavior in a simple periodic market display a fundamental behavior where performance tracks effort, aspirations track performance, and effort tracks aspirations. The simulation illustrates how the evaluation of performance leads to adjustments in performance targets (so-called "floating goals"). At the same time, the evaluation determines organizational action, and thus also future performance. These two feedback loops have opposing effects. Performance above aspirations leads to positive evaluation. This gives less effort and lower performance. At the same time, positive evaluation also leads to increased aspirations. This gives lower evaluation, increased effort and higher performance. The result is a system which oscillates with increasing amplitude in an unstable manner. However, the limitation on organizational energy stabilizes the system, which oscillates between high and low levels of effort, performance and aspirations. There does not appear to be any type or magnitude of periodic environmental input that does not give oscillating response.
In this initial simulation we have assumed well behaved market input, perfect monitoring of competitors, and rational organizational behavior where action influences performance in a deterministic manner. We modify the model parameters to describe partly chaotic input from the market, imperfect monitoring of competitor performance, and boundedly rational organizational action. On a detailed level, we see various forms of random fluctuations and there does not appear to be any systematic relation between variation of the different variables. Phase plots between pairs of variables also illustrate this randomness in system behavior, but the principal axes in all phase plots are the same. That is, the overall phase relationships between variables are the same as before. This illustrates how the overall dynamic behavior of the system is stable with respect to randomness in environmental input and organizational behavior.
The possibilities - implications and extensions
In several areas our model is a simplified description of learning in real organizations, and we outline a set of ideas for extending and refining the model content and behavior
- Introducing various delays in order to model the effect of variable learning rates in organizations.
- Introducing variable amounts of search and slack, in order to model the effect of organizational growth and decline (different levels of effort)
- Introducing more realistic models of organizational action, such as interfacing the model with discrete event simulation of information processing in the Virtual Design Team (VDT) simulator, in order to determine the effectiveness of organizational effort as a combination of efficiency and quality measures. Another possibility is to interface the model with a Garbage Can Model, using the rate of solved problems as a measure of organizational effort.
Our work so far only covers a limited set of studies, as a start for a more thorough investigation of organizational action and learning. We have not yet compared the simulation results with performance histories from real organizations, but propose to do so in future work. Despite these simplifications and limitations, the initial simulation results illustrate that "what" and "how" organizations learn is influenced by input from the market, by the rationality of their own action, by the quality of their performance measurement systems, and by their ability to interpret the performance of their competitors. However, the details of neither affect the overall dynamics of learning systems in organizations.
The simulation results suggest that organizations should pay attention to and seek to control those elements that influence their learning: However, they should not expect that all efforts to control learning will have consistently positive effects on performance and behavior.
The results also suggest a more general "contingency principle of organizational learning" - there may be no best way to manage learning, in organizations, but learning efforts will be more effective if they are matched to the action, environment, organization and target aspiration levels of the organization.
THE EVOLUTION OF SKILL SETS IN CONSULTING FIRMS
Robert N. Bernard
Coopers and Lybrand Consulting
Consulting firms typically have specialties that have evolved over time. Industries in which a specific firm specializes are highly dependent upon the skills of the firm's employees. The entire set of employee skills, the firms' skill set, is dynamic -- over time, employees learn more about their trade, they leave the firm, and new employees (with new skills) join the firm. Presumably, a firm's skill set evolves to better solve tasks that the firm is given.
I have created a computer simulation that attempts to replicate this evolutionary process. The simulation is written in C++, and implemented on a Silicon Graphics IRIX workstation. Certain tasks come into the simulated firm, consultants are grouped together to solve the task, and the tasks are solved.
The method of representing firms, employees, and tasks is straightforward. Employees are conceived of primarily a set of skills. This set of skills is represented as an ordered vector of integers, in which each element of the vector ranges from 0 to 9. The length of the vector varies depending on the job level of the employee, vectors with more elements indicating higher levels. Tasks are also conceived of as an ordered set of individual subtasks. Similar to employees, each subtask is an integer from 0 to 9. Firms are represented as (a) a set of tasks to be completed and (b) a set of employees that will complete the tasks.
Tasks are solved at a certain quality level that depends on the distance between an employee's skill set and the corresponding ordered subtasks the employee is solving. The closer in distance an element of the employee's skills is to each subtask, the higher the quality level the employee will complete the task. Depending on the quality of the solution, certain employees are promoted to higher and higher management levels while other employees are dismissed. The skill set of the firm changes to reflect the endogenous (employee driven) changes and the exogenous (firm and task driven) factors.
Two or more firms can be co-evolved together, bidding on certain tasks and -- if they win the task -- achieving some quality level while completing the task. Depending on the number of different types of tasks that are given to the firms, the skill sets that emerge within a firm are different. In other words, firms evolve (as a whole) to fill cetain tasks niches in the environment by promoting those employees that perform the best and firing those that cannot perform. Thus, over time, a certain firm may evolve to be very good at solving a certain type of task, but not good at solving any other type of task. If the types of tasks that the firm can bid upon changes drastically, the firm's ability to perform will be hampered.
Via the simulation, we can explore the adaptation of firms to the tasks they are given, the differentiation of skills within a consulting firm, the distribution of employees at different job levels, the emergence of consulting teams, and the coevolution of competing firms.
Early, preliminary results suggest that:
1. employees in the firms evolve to match the tasks given, and thus, the firm adapts overall;
2. the employees' skills distribution depends not only on the type of incoming tasks, but also on the speed of incoming tasks;
3. the distribution of simulated employee job levels corresponds well to the distribution of employee job levels at actual consulting firms, with both the simulated and actual numbers resembling a rank-size distribution;
4. certain employees tend to work together more than others, and thus, teams begin to form; and,
5. firms evolve to fill task niches in their environment.
A BAYESIAN MODEL OF PANIC IN BELIEF
Carter Butts
Carnegie Mellon University
Department of Social & Decision Sciences
One commonly observed phenomenon in the study of belief is what has been called the "consensual validation of reality": the idea that persons in highly inbred social networks alter their beliefs regarding the external world by repeated interaction with each other rather than by direct observation. In this paper, a Bayesian conditional probability model will be used to explore the conditions necessary for such outcomes, and alternative results will be likewise documented. Finally, suggestions for operationalization of the Bayesian model in experimental research will be given, along with some implications of the theory for common phenomena such as the propagation of ideas by media sources.
In his 1993 study of the satanic cult panics of the late 1980's and early 1990's, Jeffery Victor repeatedly observed the consensual validation of reality at work among panic victims. These persons were confronted by terrifying stories of mutilation and murder, tales which few if any of them were in a position to verify firsthand. Lacking direct evidence, they naturally turned to those around them for information and advice; in many cases, however, the givers of advice did not themselves have any sources of information besides other panic victims. As a result, once the panic was underway frightened citizens encountered nearly identical accounts from multiple, trusted sources, further cementing the belief that they were under siege from diabolical, bloodthirsty cultists who sought to kidnap and sacrifice their children in the name of Satan.
Needless to say, there never were any cultists, nor even any verifiable cases of ritual murders (Victor, 1993). This did not prevent, however, the panic from spreading and even from building permanent enclaves in the psychiatric and religious communities (Victor, 1993)(Victor, 1995). With no evidence to go on but the testimony of others, large numbers of persons were mislead; but this was in no way an isolated incident. Indeed, at least since Charles Mackay's 1841 study of Extraordinary Popular Delusions and the Madness of Crowds there has been an awareness of the volatility of popular opinion, and the literature on moral panics and herd behavior (both popular and scholarly; see Bulgatz (1992) for an amusing and informative example) is positively littered with examples of such collective action.
An overview of the work on moral panics in the sociological tradition can be found in Goode and Ben-Yehuda (1994). This perspective has tended to see moral panics as resulting from interactions between grassroots concerns and elite interests (Goode and Ben-Yehuda, 1994) but does not offer a formalized framework from which to make predictions. Recent interdisciplinary work, however, has been able to make progress in building quantitative models of collective behavior, particularly those which capture the essential role of information in the process of creating "aggregate results which the individual neither intends, nor needs to be aware of" (Schelling, 1969: 488). While some models of collective phenomena (such as Schelling's segregation model (Schelling, 1969), or Coleman's control model (in Coleman, 1990: 197-240)) do not explicitly consider imperfect information, even these models nevertheless tend to depend on imperfect or incomplete information in order to function. The new focus on information can be seen in the use of so-called "herd behavior" models in the economic literature (see Shiller, 1995). Abhijit Banerjee (1992) and Bikhchandani et al. (1992) describe "informational cascade" effects when a first-mover's (possibly arbitrary or noisy) choices act as signals to later movers; others, such as Lux (1995), treat information diffusion and group influences leading to long-term price inefficiencies. All of these treatments are valuable and informative, but work still remains. As Schiller points out, the information cascade phenomenon is limited in applicability (Schiller, 1995: 185). Likewise, Lux's contagion model (though perfectly adequate for his purposes) is founded at the mezzo level and does not consider the interplay of information and individual belief implicated in the social psychological accounts such as those of Asch (1952), Deutsch and Gerard (1955), and Sherif (1936). In an effort, then, to move towards a micro-level model of the dynamics of belief, we will here consider a simple model based on subjective probabilities which yields behaviors consistent with the above mentioned empirical accounts of contagious belief and the consensual validation of reality.
The core of the Bayesian belief model is a learning model. This model is, in essence, a very simple and direct application of Bayes' rule to a hypothetical situation in which an actor observes another actor's "belief" or "disbelief" signal in reference to some possible event. For purposes of this investigation, we presume a minimal case in which two actors interact with each other exclusively and repeatedly over time, in which signals are informative, and in which each actor regards the other with the same level of credibility. Though interactions are here seen as purely dyadic, repeated, and exclusive, this is not a requisite condition for many of the model's results.
A phase portrait analysis and numerical simulation of the belief model allows us to make some clear predictions regarding actor belief changes in highly inbred social networks. The first, strongest implication of the behavior of the belief model is that those with similar beliefs should reinforce each other with extreme rapidity. Two persons, for instance, who were barely more than indifferent towards the idea that a hypothetical event is true would, for a wide range of parameters, be wholly convinced were they to interact with each other a mere three or so times. This prediction may be generalized, with some restrictions, to the claim that individuals in highly inbred groups with diverse but essentially allied beliefs will tend rapidly to alter their beliefs to a more extreme form, that this shift will be group-wide, and that its speed will not depend appreciably on the size of the group. Further examination of the model suggests that, under certain conditions, the polarization process will "favor" belief or disbelief outcomes; in some cases, this may allow for "invasion" of a population by an opposing belief state.
The above discussion considers groups of actors, such as cliques, with equivalent mutual interactions. Another application of the model's results is in the consideration of the effect of a single "broadcast" source on a population of actors. While the precise dynamics produced by such a phenomenon depend on a number of factors, it is suggested that these sources may be able to "touch off" panics among appropriately structured groups. By relaxing some of the model's basic assumptions, predictions for these and other, more complicated situations may be produced, allowing for empirical tests in a wide range of circumstances. It is hoped, then, that this Bayesian model of belief will help to guide future research into the nature of panic behavior.
REFERENCES:
Asch, Solomon. Social Psychology. Prentice Hall, NJ, 1952.
Banerjee, Abhijit V. "A Simple Model of Herd Behavior." Quarterly Journal of Economics; Vol 107, No 3, August 1992, pp797-817.
Bikhchandani, Sushil; Hirshleifer, David and Welch, Ivo. "A Theory of Fashion, Custom, and Cultural Change." Journal of Political Economy; Vol 100, No 5, October 1992, pp992-1026.
Bulgatz, Joseph. Ponzi Schemes, Invaders from Mars and More Extraordinary Popular Delusions. Harmony Books, New York, 1992.
Coleman, James. Foundations of Social Theory. Harvard University Press, Cambridge, 1990.
Deutsch, Morton and Gerard, Harold B. "A Study of Normative and Informational Social Influences upon Individual Judgment." Journal of Abnormal and Social Psychology, Vol 51, No 3, November 1955, pp629-36.
Goode, Erich and Ben-Yehuda, Nachman. "Moral Panics: Culture, Politics, and Social Construction." Annual Review of Sociology, Vol 20, 1994, pp149-71.
Lux, Thomas. "Herd Behaviour, Bubbles and Crashes." The Economic Journal; Vol 105, No 431, July 1995, pp881-96.
Schelling, Thomas C. "Models of Segregation." American Economic Review; Vol 59, No 2, May 1969, pp483-93.
Sherif, M. The Psychology of Social Norms. Harper, New York, 1936.
Shiller, Robert J. "Conversation, Information, and Herd Behavior." American Economic Review; Vol 85, No 2, May 1995, pp181-85.
Victor, Jeffrey S. Satanic Panic: The Creation of a Contemporary Legend.
Open Court, 1993. -- "The Dangers of Moral Panics." Skeptic; Vol 3, No 3, 1995, pp44-51.
A SIMULATION OF A NETWORK OF AGENTS SOLVING TRAVELLING SALESMAN'S PROBLEMS
Kevin Crowston
Syracuse University
School of Information Studies
and
Bernardo Huberman
Xerox Parc
1. Introduction
Informal communication networks are clearly important to organizational performance. We would therefore like to know what patterns of communication contribute the most under what circumstances, as well as how general and robust these patterns are. The purpose of this study is to develop a computer simulation of an organizational communications network and explore how it relates to the organizations problem solving performance.
To answer these questions, Huberman and Hogg (1995) presented a model of the network of informal communications within an organization composed of multiple actors. They assumed that actors in an organization improve their performance (and thus the performance of the whole organization) by getting hints from other actors. They developed a simple model of the effects of hints from which they calculated what pattern of communications help most given the size and diversity of the organization. Their model suggested that small uniform organizations have the best performance when they are flat (i.e., everyone talks about equally to everyone else) but as size and diversity increase, performance is improved by clustering (i.e., actors get hints primarily from a subset of the other actors). Using a very general model of actor learning, they analyzed the stability of patterns of interaction and concluded that flat organizations become unstable more readily than clustered ones.
2. My model
Huberman and Hogg (1995) argued that the phenomena they modelled were general. Their conclusions, however, were based on a probabilistic analysis of the distribution of hints and their utilities. In this paper, I will discuss a computational model that attempts to replicate some of the properties of the networks studied by Huberman and Hogg, but where computational actors cooperate to solve a real problem. As with Huberman and Hogg's model, actors can provide hints, such as decompositions of or partial solutions to the problem. Sometimes a hint will be useful (i.e., help the recipient solve its problem more quickly); in other cases, it may not help or may even be incomprehensible to the recipient. As in Huberman and Hogg's model, actors keep track of the utility of hints and adjust the probability of interaction accordingly: the probability of asking another actor for a hint will increase if that actor's hints have been useful in the past and decrease if not.
In the remainder of this section, I will discuss the reasons for our choice of problem and describe the inputs, operation and outputs of the simulation.
2.1 The problem
The actors in the simulation work cooperatively on a traveling salesman problem (TSP). The input to the TSP is a list of nodes and a matrix of distances between the nodes; the goal is to find the minimum length tour that visits each node exactly once and returns to the origin. TSP was chosen because it is a hard (NP-complete) problem that can be solved by computer agents and because there are numerous sample problems and solution techniques available. Importantly for this simulation, information from one approach can be useful for others, suggesting that exchanging hints will be useful. For example, a current best tour can be used as input to a branch-and-bound or a path improvement algorithm.
A potentially important difference between our work and Huberman and Hogg's is that they assume performance is additive (i.e., the group performance is the sum of group members' performances), while for this task, the group performance equals the performance of the best member of the group (although individual performance depends on what hints they get, so it's not quite so simple). It's not clear how or if this difference will affect the outcome of the study.
2.2 Inputs
There are two inputs to the simulation: the organization and the problem to be solved. For the organization, we input the type and number of actors . A particular organization can be made homogeneous or diverse, larger or smaller, simply by selecting the actors which compose it. We can also supply the pattern of interactions between the actors or allow them to be choosen randomly. The problem is drawn from TSPLIB, a collection of sample TSPs of varying sizes. If the length of the optimal solution is known, it is used to stop the simulation when the solutions found are close enough to optimal.
2.3 How the simulation works
The simulation is a discrete event simulation, where the events are sending and receiving hints. The simulation is started by reading the problem to be solved and sending it as a hint to all agents. Each agent receives the hint, determines if it is one it is one it wants to work on (as discussed below) and works on it if it is.After it works (or decides not to work) on the hint, the actor randomly selects another actor to ask for a hint, based on their interaction strengths (as discussed below). The interaction strengths may allow the actor a chance to pick itself, in which case it performs what Huberman and Hogg refer to as self-work, building on its own previous results. When the requested hint arrives, the process repeats.
Currently there are two kinds of hints: problems and solved tours. A problem is a set of nodes that need to be connected; a solved tour is an ordered set of nodes and the length of the tour that visits those nodes in that order. A hint might include only some of the nodes in the initial problem, i.e., it might be a partial tour or problem decomposition. Particular actors understand one or the other or sometimes both kinds of hints, as documented below. If the actor understands the hint, it will work on it; otherwise it just asks for another hint.
Actors rate the quality of the hints that they receive. A hint that can not be used is rated a zero; other hints are scored from 0 to 1, based on how useful they are for the particular actor. Some actors simply assign quality 1 to any hint they understand and 0 to others; others assign a quality based on the length of the partial tour. These ratings are used to update the interaction strengths. Currently the interaction strength between two agents is calculated as the sum of the scores of all hints the first agent received from the second, divided by the total number of hints received. Actors that consistently provide unusable hints will therefore have interaction strengths that approach zero, meaning that that actor will no longer be asked for hints.
When the actor works on a hint, it creates a new hint, which it stores and can provide as a hint to another agent when asked. Current actors simply provide the most recent hint they have generated.
Actors run asynchronously to avoid artifacts due to synchronized actions, such as noted by Huberman and Glance (1993). Except for the initial posting of the initial problem, all communications events are scheduled independently. For example, to request a hint, an actor calculates how long it spent working on the current hint, adds the time required to transport a request message and schedules the request to arrive after that delay. Similarly, actors asked for hints schedule the hint to arrive after the delay required to transport a hint. Currently actors use their actual run time to determine the appropriate delays. This approach is in some sense most realistic, but may make runs of the model difficult to reproduce.
2.4 Outputs
For each run, we report the best solution found and the clock time and number of messages it took to find that solution. (One solution is better than another if it includes more nodes or if it includes the same number of nodes and is shorter.) To simplify the initial experiments, we consider the problem solved when a solution is found within a fixed percentage of the known optimal solution and use only time as the figure of merit. As well, we output the interaction strengths, which can be used to initialize another run of the simulation, allowing us to rerun the simulation with the evolved interaction pattern.
3. Hypotheses
Based on Huberman and Hogg's (1995) work, we developed several hypotheses for the performance of the simulated actors.
3.1 Fit between organization composition and interaction pattern
First, using the simulation, we can study how over the course of the simulation, a pattern of interaction between actors will develop and test their predictions about the fit between population characteristics and interaction patterns. For example, in a small uniform population, we expect that a flat organization will do best, while in a large diverse population, we expect a clustered organization to do best. To test these hypotheses we can compare the performance of organizations with flat and clustered interaction strength matrices and different combinations of agents.
As well, we can look at the pattern of interactions that naturally evolves. For a large diverse population, for example, we can see if an initial flat interaction strength matrix evolves to a clustered one. We can also study how the interaction pattern evolves as the organization or environment changes.
3.2 Effect of interaction pattern on performance
Second, we hypothesize that having developed this pattern will allow the organization to solve new problems more quickly. We therefore measure time to solve with a random or flat interaction matrix and compare this a previously evolved interaction pattern. We will compare the performance of organizations with evolved and random interaction matrix on TSP problems other than the only used to develop the matrix.
4. Implementation
The simulation was written in Objective-C using the Swarm simulation environment. Swarm provides a discrete-event simulation framework, as well as probes, controls and other graphical interface features. An earlier version of the simulation was written in C and used Sun's threads package to implement concurrent execution, but this version was abandoned due to problems with the threads and the attraction of the Swarm environment. However, the initial results reported here are taken from that prototype.
4.1 The actors
The list of actors implented is shown in Table 1. To ensure a diversity of hints, some of these actors implement all or part of what are believed to be reasonable algorithms for solving TSPs while others are very simple-minded approaches to the problem. Several of the actors were adapted from source developed by R. J. Craig of AT&T Bell Labs, Naperville, IL (kat3@uscbu.ih.att.com).
5. Preliminary results
I developed four different organizations, large (20? actors) vs. small (6 actors?) and diverse vs. homogeneous (homo). These organizations were used to solve two problems from TSPLIB, bayg29 and pr226, with 29 and 226 points respectively. The simulation was run until the solution was within 2% of optimal for problem bayg29 and within 10% for pr226. I ran the simulation 11 times for each condition, using a flat interaction matrix for the first run and then using the evolved interaction matrix from one run as the starting point for the next. Unfortunately the running time does not clearly decrease on successive runs, as predicated. Instead, the data are quite noisy, making conclusions difficult.
Table 2 shows the average quality of the solution and running time for the different conditions. Note that the homogeneous organizations took much longer to find a solution than did the diverse organizations. Also, some of the qualities are greater than 110%, because the current stopping rule relaxes the quality target as the run time increase. For the larger problem, the large organization was better of the diverse ones while the small one was the better of the homogenenous. For the small problem, the smaller organizations took less time for both conditions.
6. Things left to be done
Clearly more runs are needed to determine statistically reliable results.
6.1 Learning and memory
Currently, agents learn which other agents provide usable hints but in a very unsophisticated way. I would like to draw on other work in multi-agent learning to enhance this aspect of the simulation.
As well, current agents only remember their most recent result. The performance might be altered if agents remembered a selection of past results.
One extension I've already implemented is to allow actors set an "aspiration level", and not work on hints of lower quality than they've created. Without this kind of memory, run times were exponential in the length of the hint-agents would lose a long path when given a shorter hint, making it harder and harder to build long paths. However, the current version is too locked in to the current results. A better solution might be to make the aspiration level more flexible, to avoid a lot of backsliding, but allow more useful variation in the results.
6.2 Symbols
To investigate phenomena related to legitimate peripheral participation, I plan to make some hints symbolic. Parts of problems or tours will be encode symbolically. These messages would be understandable only to other agents that understand the meaning of those symbols.
6.3 Additional agent types
Table 3 presents additional actors that it would be interesting to implement. Some of these algorithms provides a useful lower bound to the problem, but require a third type of hint, namely, a 1-tree.
As well, it would be interesting to develop "organizational" actors, such as brokers, intermediaries, etc. For example, we might add memory to the organization in the form of an actor that simply remembers the best hint it's ever seen.
REFERENCES
Clearwater, S., Huberman, B. A. and Hogg, T. (1991). Cooperative agents?
de Souza, P. S. (1993). Asynchronous Organizations for Multi-Algorithm Problems. Doctoral Thesis, Department of Electrical and Computer Engineering, Carnegie-Mellon University.
Huberman, B. A. (1990). The performance of cooperative processes. Physica D, 42, 38-47.
Huberman, B. A. and Glance, N. S. (1993). Evolutionary games and computer simulations. Proceedings of the National Academy of Science, USA, 90(August), 7716-7718.
Huberman, B. A. and Hogg, T. (1995). Communities of practice: Performance and evolution. Computational and Mathematical Organization Theory, 1(1).
Lin, S. and Kernighan, B. K. (1973). An effective heuristic algorithm for the traveling-salesman problem. Operations Research, 21(2), 498-516.
Norman, B. A. and Bean, J. C. (1994). Random keys genetic algorithm for job shop scheduling (Technical Report 94-5). Department of Industrial and Operations Engineering, University of Michigan.
Table 1. Currently implemented actors.
Actor name Algorithm implemented
calc-len Given a problem, constructs a tour by connecting the nodes in the
order in which they appear in the problem
ni Given a problem, constructs tours by nearest insertion (find the
point nearest the last one inserted and insert it where it adds the
least length to the tour)
fi Given a problem, constructs tours by farthest insertion (insert the
point farthest from the last inserted)
ai Given a problem, constructs tours by arbitrary insertion (inserts
the points in the order they appear in the problem)
simple- Given a problem, decomposes it into 3 node divider subproblems based
on the order the nodes appear in the problem
cluster- Given a problem, decomposes it into divider subproblems of length 3
by picking a node and the 2 other closest nodes
join3 Given a 2 or 3 node problem, constructs a tour with the nodes in the
optimal order simple-joiner
Given two non-overlapping tours, constructs a tour that links them in
the optimal order
twoopt Given a tour, tries to shorten it by deleting two of the links and
recombining the segments in all possible ways
threeop Given a tour, tries to shorten it by deleting three of the links and
recombining the segments in all possible ways
Table 2. Average running time by condition.
Problem
Org size Org comp Results bayg29 pr226
large diverse Average 102% 111%
solution
Average time 2.44 353.02
homo Average 101% 113%
solution
Average time 23.65 5856.97
Average 101% 112%
solution
Average time 13.55 3104.99
small diverse Average 102% 106%
solution
Average time 2.20 920.83
homo Average 102% 113%
solution
Average time 12.01 1287.06
Average 102% 110%
solution
Average time 7.11 1103.94
Table 3. Additional actor types to be implemented.
Actor name Algorithm implemented
*or-opt Given a tour, looks for a shorter tour by moving segments of 1, 2 or
3 nodes to another position in the tour
*rand-4opt Given a tour, randomly deletes 4 edges and reconnects the segments
(part of iterated LK (Lin and Kernighan, 1973, p. 38))
*linkern Given a tour, improves it by running one pass of the Lin-Kernighan
algorithm (Lin and Kernighan, 1973)
*pmx Given two tours including the same nodes, breeds them using
partially-mapped crossover
*bean Given two tours including the same nodes, breeds them using random
keys (Norman and Bean, 1994)
*mixer Given two tours including the same nodes, creates a new tour using as
far as possible only edges from the tours
*bnb Given a tour, searches for a tour using branch and bound; uses solved
tours to limit search
*?heldkarp Given a problem, finds a lower-bound by finding a non-feasible 1-tree
*?tree-mixer Given a tour and a 1-tree, mixes them following the mixer algorithm
PROMOTING OR DISCOURAGING CONFLICT IN ENGINEERING PROJECT TEAMS: INSIGHTS FROM A COMPUTATIONAL ORGANIZATIONAL DESIGN PERSPECTIVE
Raymond E. Levitt, Jan Thomsen, and Yul K. Kwon
Stanford University
One of the oldest problems in the field of organizational theory and management is the challenge of organizing behavior within groups whose members possess discrepant goals and priorities. In project teams, disagreements between participants typically arise when two or more actors favor the adoption of different solutions with which to meet requirements. The cause of such disagreements is twofold. First, actors may lack sufficient knowledge of the constraints that bear on the solution at the time that they select solution alternatives, forcing them to make choices in light of incomplete information. Second, the preferences and beliefs of each individual may result in actors giving different weights to the various criteria by which each alternative is evaluated, causing them to rank alternatives differently. We refer to these task conflicts as problems of goal incongruency.
Background
Organizations use a number of different mechanisms to minimize the effects of goal in congruency and to provide managers with control over the behavior of subordinates. These mechanisms, which include the monitoring of employees, as well as their selection, training, and socialization, have been widely discussed in the literature on management science and neoclassical economics, as interest in the internal processes and dynamics of organizations has increased among researchers over the past several decades (Ouchi, 1979; Eisenhardt, 1985, 1989; Levinthal, 1988; Milgrom & Roberts, 1992). The core assumption underlying many of the conventional theories of goal in congruency is that any degree of goal incongruency will be detrimental to project team performance. This view is based on studies demonstrating that deviation from managerially-prescribed goals by subordinates will necessitate additional coordination and communication efforts to resolve the discrepancy. Such theories, which consider organizations as giant "Weber machines," generally maintain that conflicts between organizational participants invariably decrease organizational efficiency.
The problem of goal incongruency is exacerbated in large engineering projects due to the sheer complexity of modern engineering artifacts. Modern engineering projects often require actors from different areas (e.g., disciplines, departments, subcontractors) to coordinate with one another in order to accomplish the overall objective successfully. The need for high levels of interaction between diverse groups prohibits organizations from simply decomposing tasks and responsibilities and assigning them to strictly delineated departments. Consequently, not only must organizations deal effectively with goal incongruency problems arising within supervisor-subordinate relationships, but they must also negotiate goal incongruency problems arising in lateral relationships between peers working on interdependent activities.
Traditional views of goal incongruency in engineering projects have recently been challenged by new research within the field of social psychology, indicating some potential positive benefits of goal incongruency on organizational performance (e.g., Amason, 1996; Jehn, 1995; Pelled, 1996; Watson, et al, 1993). On a macro-organizational level, these theories maintain that goal incongruency can increase the diversity of an organization's behavioral repertoire for meeting the demands imposed by a dynamic environment. In other words, the absence of any goal incongruency has the potential to compromise an organization's capacity for adaptation by reducing the range of available behavioral patterns with which an organization can respond to changes in the environment (Weick, 1979). On a micro-organizational level, goal incongruency is theorized to confer two distinct advantages. It will force actors to consider a wider range of possible solutions to a problem, which increases the likelihood that a more ideal solution will be found. Moreover, it will lead to a greater understanding and clarification of the trade-offs associated with each solution under consideration, and will encourage actors to formalize their knowledge of these trade-offs implicitly or explicitly into a "goal trade-off table." Shared goal trade-off tables among project participants can be viewed as a common set of values or culture. The existence of shared values and culture is now widely viewed to increase efficiency by serving as a set of guideposts that allow actors to make decisions more quickly and consistently when similar problems arise further downstream.
To date, research in management science and in economic theories of the firm has been unable to analyze the emergent effects of goal incongruency at a sufficiently detailed level of granularity to generate practical guidance for managers. The chief obstacle hindering the translation of theoretical knowledge into practical knowledge is that most social psychological theories of agent models with goal incongruency were developed from short-lived experiments involving dyadic relationships between two people, rather than long-term experiments involving larger groups of people. Consequently, little data exists on how goal incongruencies among large groups of individuals can affect the performance of organizations as a whole, or how goal incongruencies interact with other organizational design factors, such as the level of centralization, monitoring, and matrix strength. Insofar as organizational behavior is an emergent phenomenon, and the effects of goal incongruency on organizational behavior cannot be predicted based on an understanding of its effects on a single relationship, we believe that the most promising approach to analyzing goal incongruency in organizations comes from the field of computational organizational theory. Therefore, we have chosen to use a computational model of organizations to simulate organizational behavior at the micro-level in order to analyze the relationships between goal incongruency and organizational performance.
The Virtual Team Alliance (VTA) Computational Model
Many variables must be considered in modeling the emergent effects of goal incongruency on organizational behavior arising from the interaction of multiple actors. We chose to implement our framework through a probabilistic discrete-event simulation model. Specifically, we decided to ground our models the Virtual Design Team framework (VDT) (Jin & Levitt, 1996) rather than other modeling frameworks such as the garbage can model and its derivatives (Cohen, et al, 1972; Masuch & Lapotin, 1989) because of the detailed grain size of actor and task modeling afforded by the VDT platform, the specificity of its measures for performance, and its theoretical grounding in the information processing view of organizations. We extended the existing VDT platform substantially in order to enlarge its expressiveness to describe our problem domain.
The extended VDT model, which we call the Virtual Team Alliances (VTA) model, is a new tool for investigating the emergent effects of goal incongruencies between individual or group actors on project team performance. In our model, a manager may respond to goal incongruency between himself or herself and a subordinate by investing in additional monitoring efforts, or by delegating less decision-making authority to the subordinate. Actors will respond to goal incongruency with another task-interdependent actor by engaging in one of four possible behaviors:
(1)searching for alternatives, (2)clarifying goals, (3)steamrolling, or
(4)politicking.
Our presentation will describe in more detail how we incorporate and operationalize each of these behaviors into our framework based on well-
established social psychological and economic theories of micro-organizational behavior. We will also elaborate on how our framework has been applied to two on-going aerospace engineering projects for designing launch vehicles, and present final results of the experiments that were conducted using VTA models of these projects.
Results of Preliminary VTA "Virtual Experiments"
Our limited suite of "virtual experiments" conducted to date with the VTA model used a two-factorial design. Preliminary results indicate that project performance can vary significantly for project teams with both the level of goal incongruency and the level of monitoring. The two project teams we selected differed from one another in regard to a number of major project parameters, including the levels of goal incongruency, formalization, centralization, and matrix strength. Our experiments indicated that neither extremes of goal incongruency would be conducive to maximizing project performance in these two projects; and results suggest that the optimal level of goal incongruency is influenced by organizational differences between the projects.
The first project team, which had higher centralization and formalization levels and lower matrix strength, performed best (in terms of duration, cost and quality) at a level of goal incongruency lower than that of the second project. Moreover, the greatest gains in project performance were found to occur when monitoring was increased in the first project, and when goal incongruency was decreased in the second project. The second project team experienced severe bottlenecks in its work process during the design phase. By changing the level of goal incongruency and the level of monitoring, our experiments determined that these bottlenecks could be greatly reduced.
Outline of Proposed CMOT Presentation
In our CMOT talk, the first author will present the VTA framework, explain the design and analysis of preliminary VTA experiments, and conclude with a discussion of the significance of this research for both organizational theorists and project managers.
REFERENCES:
Amason, A. 1996. "Distinguishing the effects of functional and dysfunctional conflict on strategic decision making: Resolving a paradox for top management teams." Academy of Management Journal, 39: 123-
148.
Cohen, M. D., March, J. G. & Olsen, J. P. (1972). "A Garbage Can Model of Organizational Choice." Administrative Science Quarterly, 1.
Eisenhardt, K. M. 1985. "Control: Organizational and Economic Approaches." Management Science, 2:134-149.
Eisenhardt, K.M. 1989. "Agency Theory: An Assessment and Review." Academy of Mangement Review, 1:57-74.
Jehn, K. 1995. "A multimethod examination of the benefits and detriments of intragroup conflict." Administrative Science Quarterly, 40: 256-282.
Jin, Y. & Levitt, R. E. 1996. "The Virtual Design Team: A Computational model of Project Organizations," Computational and Mathematical Organizational Theory, 2.2.
Levintal, D. 1988. "A Survey of Agency Models of Organizations, "Journal of Economic Behavior and Organization, 9:153-185.
Masuch, M. & LaPotin, P. (1989). "Beyond Garbage Cans: An AI Model of
Organizational Choice." Administrative Science Quarterly, 1:38-67.
Milgrom, P & Roberts, J. (1992). Economics, Organization & Management. Prentice-Hall Inc.
Ouchi, W. 1979. "A Conceptual Framework for the Design of Organization Control Mechanisms." Management Science, 25:833-848.
Pelled, L.1996. "Demographic diversity, conflict, and work group outcomes: An intervening process theory." Organization Science, 7:615-631.
Watson, W., Kumar, K & Michaelson, L. 1993. "Cultural diversity's impact on interaction process and performance: Comparing homogeneous and diverse task groups." Academy of Management Journal, 36:590-602.
Weick, K. E. (1979). The Social Psychology of Organizing. McGraw-Hill Inc.
MODELING MEDICAL PROCESSES FOR COMPUTATIONAL ORGANIZATION SIMULATION: LINKING PROTOCOL DESCRIPTIONS TO SIMULATION REQUIREMENTS
Douglas B. Fridsma, M.D
Section on Medical Informatics
Stanford University School of Medicine
and
Jan Thomsen
Construction Engineering and Management
Stanford University School of Engineering
Motivation
Computer simulation of organizations requires formal, computable representations of real world objects. Creating formal representations of people, organizations and work processes for simulation require substantial modeling ef-fort, and there would be an advantage to re-using information from existing "best-practice" organization or work process descriptions. In medicine, medical protocols for patient care represent the "best practice" work processes for a given patient condition and are a natural starting point for creating organizational simulations. These protocols are often drawn from national consensus committees, and thus do not contain information specific to an organization=97they assume an ideal, uncapacitated organization in which all activities are completed without exception, all patients respond in categorical ways, and all interactions within the organization are seamlessly coordinated. For a medical protocol to be used as part of an organization simulation, the medical protocol alone would be inadequate. Without a model of the coordination, communication, and exception handling requirements of a particular protocol, the simulation of this protocol within an organization would be inaccurate at predicting the performance characteristics of the protocol.
If we wish to use these medical protocols as the starting point of our model, we will need to modify these protocols for the requirements of the simulation. It is unreasonable to expect the developers of national protocols to include all the information that is necessary to simulate a protocol, and it is impossible for the developers of these protocols to know the specifics of each organization in which the protocol might be used. We describe a methodology that draws on the artificial intelligence (AI) literature, organization literature and engineering management methods to derive these protocol parameters in a structured way. Our paper describes the necessary inputs into our computer simula-tion, and how we can derive these input parameters from a medical protocol using protocol goal abstraction, match-ing these abstractions to domain and organization knowledge, and refining the parameters of the domain ontology to fit the simulation requirements.
Our Simulation Requirements
For our simulation, we have chosen an information processing view of organizations. This fits well with the medical domain since medicine is information-intensive in practice, and we desire to derive a more explicit representation for the coordination and communication burden that protocols can place on organizations.
For a realistic information processing simulation, the coordination processes needed have been described previously [e.g., Jin & Levitt, 1996]. First, we need to estimate the need for communication. This estimation is derived from the uncertainty in the activity, and the information interdependence between activities [Thompson, 1967; Galbraith, 1973]. The second parameter is derived from Simon's work on bounded rationality-exception generation [Simon, 1957]. The number of exceptions generated will be determined by the complexity and flexibility of the task (the more complex the task, the more likely an exception will arise), and the exception interdependence between these activities (which activities are affected by the exception). If we wish to create a simulation based on these two principles in organization theory, we will need to derive from a medical protocol, estimations of complexity, uncertainty, flexibility, and interdependence.
Modeling Framework: Heuristic Work Process Abstraction and Refinement
We have drawn on previous work in knowledge representation in the AI community to organize the way in which our model is developed [Clancey, 1985]. In many ways, creating a work-process model for the purpose of simulation is not unlike creating a knowledge base for inference. An expert system has an internal representation of the world whose behavior mimics that of real world experts. We need to create a computer model of a particular protocol that has the characteristics of the real world protocol that it represents. Heuristic classification is an abstract method of computation that has been the foundation of many expert systems and planning systems [Clancey, 1985]. It requires abstraction of a series of data to more general concepts, matching these general concepts to a set of possible solutions, and then refining the solution until the most specific solution is found. Although this framework is principally used to describe expert system problem solving at the knowledge level [Newell, 1982], its general framework has similarities to the tasks required to generating input to simulation models.
Figure 1 below shows our framework for creating the simulation model from the description of the protocol. The principle components of the model are shown in the grey boxes. The goal decomposition is derived from the protocol description by taking the high level goal of the protocol, and creating a decomposition of intermediate goals. This goal decomposition describes the protocol in terms of part-of relationships, in the same way a task decomposi-tion or requirement-solution diagram describes other processes.
The goals in the goal decomposition are then matched to organization-specific domain information in the domain ontology. The domain ontology describes the same activities and goals that are present in the goal decomposition, but are organized in a classification tree in which similar activities are grouped together according to the goals that they achieve. Because this description is specific to the organization being studied, it contains more specific details about the activities within the organization, and how these activities can provide implementation solutions for the protocol. The process input parameters to our simulation-complexity, flexibility, uncertainty, and interdependence--are derived from both the goal abstraction and parameter refinement. These are shown in italics in the diagram below.
We will describe each of these steps in the subsequent sections.
Goal Abstraction and Decomposition
We begin with a protocol that has been developed by a consensus committee, and which describes in detail the best practice for managing a particular patient condition. Our goal is to generate the input parameters necessary for simulation, using the protocol as a starting point. In most circumstances, the protocol will describe the purpose for using the protocol (the overall goal), and the activities that need to be done in order to achieve that goal.
First, we need to create a decomposition of the goals of the protocol. A medical protocol is generally created for a specific high-level purpose-the protocol may be used for giving a particular medication, or for the broader purpose of treating a type of cancer. These broader goals are usually well specified in the protocol, although the intermediate goals are often not explicit. Using the protocol as a guide, experts can enumerate intermediate goals in the same way that means-ends chains or a task decomposition can be developed.
For example, a protocol designed to "treat lung cancer" has a series of subgoals: "determine disease stage", "administer chemotherapy", and so forth. Each of these subgoals can be broken down into further goals: "determine presence of bone metastasis", or "manage chemotherapy side effects" for example. This process continues until suf-ficient detail is achieved in the goals such that activities can be easily associated with each of the goals. We continued the decomposition until the goals were sufficiently detailed such that they could be associated with only 2 or 3 protocol activities.
Once we have developed the goal decomposition, we must association each of the specific protocol activities with their broader goals listed in the goal decomposition. For example, the activities "give etoposide chemotherapy" and "give cyclophosphamide" (two types of chemotherapy drugs), contribute to the goal of "administering chemotherapy". Even though there may not be an explicit link between these two drugs in the original protocol, we know that an exception or complication of giving one of the chemotherapy drugs will have an impact on the overall success of the goal "administering chemotherapy", and an exception in administering one drug may require a modification in the doses of the other chemotherapy drugs. The goal decomposition provides an way of making these implicit links between activities more explicit.
Once we have determined the connections between goals and protocol activities, we can define activity complexity as the sum of goals to which that activity contributes, modified by the number of possible solutions that need to be considered [Thompson, 1967]. Our goal decomposition gives us the connection between activities and goals, and the domain ontology (described below) provides a description of alternative solutions to that goal. We can also derive activity interdependence in the same way. The protocol describes sequential interdependencies explicitly, but reciprocal interdependencies are determined from the number of shared goals between activities, modified by com-plexity of those goals.
Domain Match and Parameter Refinement
Although the goal decomposition requires some knowledge of the protocol goals to develop a reasonable model, the goal decomposition only describes attributes of the protocol. There is no information regarding how to actually do these activities, what alternative activities might achieve the same goals, or the resources that might be required to achieve these goals. The goal decomposition also fails to indicate alternatives or other related activities that might be part of problem-solving should an exception arise. For example, in our simulation, we are interested in the number of different ways in which the same activity might be completed. Activities that have a great deal of flexibility in the way in which they can be implemented require more diligent problem solving to consider all of the alternatives. Similarly, activities in which there is only one way to do things, require no problem-solving-the answer is clear.
The knowledge in the domain ontology is specific to the organization and the domain of interest (in this case, medicine). Information in the domain ontology is organized as an is-a hierarchy-a classification of activities organized by those that have similar goals. For example, the class chemotherapy administration, may have "etoposide", "cyclophosphamide", and other chemotherapeutic agents as subclasses. Etoposide can be given in a variety of different ways as well, and so additional subclasses of etoposide such as "rapid infusion of etoposide", "4 hour infusion", "continuous infusion", may be present as alternatives in the domain ontology. Each of these alternatives may have different resources requirements, different skills and other features that will affect the simulation performance.
The separation of the protocol description and task decomposition from the classification of the activities within the organization provides two advantages. First, it allows us to examine alternatives that achieve a similar goal. For some simulations, these alternatives represent the flexibility that the organization might have in implementing a goal. With many alternatives, the problem-solving becomes more difficult, and this will affect the communication and coordination requirements.
Second, the domain ontology allow us to list only those activities that the organization is capable of doing, and explicitly represent the details of the organization activities. The organization description may be relatively constant, and can be stored and maintained separately from the protocol description. By separating the protocol description from the organization description it should them be easy to run many different protocol simulations on the same or-ganization structure, since the activities can be described once, and re-used for each new simulation.
We now can derive the input parameters flexibility and uncertainty. Flexibility is simply the number of alternative ways in which a goal might be achieved. Uncertainty is related to both the number of shared goals, and the number of alternatives that need to be considered in achieving that goal.
Finally, in order to derive the input parameters of uncertainty, flexibility, interdependence, and complexity more formally, we have developed a matrix representation for these values. This makes it relatively easy to link the ab-straction and decomposition steps described above, to the simulation input parameters. Ultimately, we anticipate that it will be possible to automate much of the process of model development. If protocols have their goals (or intentions) made explicit during the process of protocol development, less manual effort will be needed to create the goal abstractions [Shahar et al, 1996]. Similarly, if organizations maintain current descriptions of the activities that they can do, separate from a particular protocol, these organization descriptions can be re-used for simulation of many different protocols.
Contributions
We have described building models for simulation in terms of abstraction of protocol goals, matching to domain and organization characteristics, and refining these descriptions to the specific input requirements of an information processing simulation. This conceptualization formalizes the tasks necessary for model building, and helps to organize the information required for simulation. This conceptualization describes the higher, knowledge level tasks that are necessary to describe a protocol for simulation We believe that this framework makes possible more automated ways of developing complex simulation models, and examining the ways in which other models are constructed.
Second, we have shown how using this framework, we can derive the input parameters for an information processing simulation. It is possible to define uncertainty, interdependence complexity, and flexibility in terms of the abstract goals in the goal decomposition and the organization alternatives in the domain ontology. Matrix representation and manipulation completes this formalism.
This method is suited to domains such as medicine in which we start with a protocol that needs to be customized for use within a simulation. Medical activities can be organized into these classification hierarchies, and the information in both the protocol and the organization re-used. As more work is done in modeling the medical domain, it should become easier to automate this process. Formalizing the process provides both a better understanding of the modeling task, and begins to make possible intelligent computer assistance with the model process.
SDML: A MULTI-AGENT LANGUAGE FOR ORGANIZATIONAL MODELLING
Scott Moss, Helen Gaylard, Steve Wallis and Bruce Edmonds
Centre for Policy Modelling
Manchester Metropolitan University
Implementors of computational models of organizations confront a trade-off between the sophistication of the representation of individual cognition on the one hand and the complexity of the modelled organization on the other. A more elaborate organizational structure is associated in the literature with a simpler representation of individuals. We introduce in this paper a programming language, SDML, which supports fast development of models and, for any representation of cognition, can entail more complex representations of organizations than we have seen previously in the literature.
SDML ("strictly declarative modelling language") has been designed to facilitate flexible multi-agent modelling of organizations. Unlike modelling architectures such as Soar or ACT-R, which incorporate specific cognitive theories in their respective agent architectures, SDML is a theory-neutral programming language. However, the requirements underlying its development, and the features whereby these are realised, mean that SDML can easily be used to represent either simple or sophisticated agents and the nature of the social relations which exist among them. Implementing models in SDML does not preclude the use of Soar's or ACT-R's particular problem-solving architectures. Because the Soar architecture in particular has been applied to a range of computational organizational models, Soar agents are being implemented in SDML.
Soar represents agents as implementations of Newell's (1990) unified theory of cognition and models written in Soar have specified at most two levels in a hierarchical organization (for example, Ye and Carley, 1995). A model in TASCCS (a "synthesis" of Double-AISS and Plural-Soar) by Verhagen and Masuch (1994) was restricted to two agents due to the implementational limitations of Soar. Carley and Svoboda (1996) represent agents by a simulated annealing algorithm or by a stochastic learning process specifically to get round the computational restrictiveness of Soar. So and Durfee (1996) represent organizations as general tree structures comprised of homogeneous agents transmitting message packets to one another.
Tambe and Rosenbloom (1996) suggest that a limitation on the complexity of computational models of agent interaction is the agent architecture. For this reason, they extend the Soar architecture to support the implementation of agents who build models of other agents whose behaviour they must track in real time knowing that they are themselves being tracked and modelled by the other agents. Tambe and Rosenbloom have not reported any models with the sort of hierarchical relationships that are essential to representing agents acting within organizations.
A number of basic requirements underlie SDML's multi-agent features. Individual agents must be able to incorporate rules determining their behaviour, including any or no cognitive theory. For these purposes a declarative representation is appropriate because it enables us to capture the distinction between behaviour and its underlying explanation. This distinction is particularly salient with respect to social phenomena of an inherently emergent nature. The ability to share rules among similar agents is convenient and this is enabled by SDML's object-oriented features. For multi-agent applications, we require the flexibility to represent such structures as organizations with arbitrarily deeply nested levels of agents. Agents must be able not only to communicate with each other but also to maintain privacy in the sense of restricted access to information. We further require the ability to represent organizations as dynamic structures which change over time. Because SDML is strictly declarative clauses once asserted to a database may not be retracted. However, change is easily modelled using time levels, an in-built feature of SDML whereby different databases are associated with agents at different periods in time.
The structure of multiple agent models is represented in SDML by a container hierarchy; for instance, persons may be contained within departments contained within firms. The outermost container is always the universe. The container hierarchy is related to a type hierarchy which supports object oriented features principally by type Composite Agent and its subtypes. The type Composite Agent allows for the representation of agent hierarchies of arbitrary depth, since any agent within a Composite Agent may itself be a Composite Agent. A type specifies the type of its container an dinherits some clause definitions via the container hierarchy. SDML's in-built predicates allow for specification in the rulebases of Composite Agents of those subagents which are active (i.e., for which the rulebases will be fired) and, in the case of Serial Composite Agents, the order in which subagents are activated.
In SDML agents normally communicate by writing the results of their rule firing to their databases or those of a container and reading the results of another agent's rule firing from that agent's database or that of a shared container. Accessibility restrictions permitting, agents can also read from or write to other agents' databases using explicit addressing. The default database to which a clause is asserted when a rule fires depends upon where that clause is defined. If it is defined in the agent's type or a supertype (i.e., inherited via that type hierarchy) then it will be asserted to the agent's own database. However, if the clause definition is inherited via the container hierarchy then it will be asserted to the database of the container where it is defined. It is often convenient for agents to share information via the container as this does not require explicit addressing. This is especially so where the structure of the organization may be changing.
Uncertainty concerning other agents' actions was mentioned above as an important feature which we require models to be able to represent if they are to be adequately expressive with respect to social phenomena. Related to this, SDML allows for different accessibility for agents to other agents' databases. For example, some decisions and actions made by a firm may never be directly accessible to other firms and, similarly, some actions of individuals within it may never be accessible to the firm. Clauses may be defined as private to the agent or public, or, intermediate between these, internal to the defining container, e.g., accessible to all agents within a firm but none outside it. There is also the facility to, for example, make an agent's clauses publicly readable but not writeable.
An example is exhibited to demonstrate how SDML's container hierarchy supports the representaion of firms containing departments which in turn contain persons. We represent the lowest-level agents as cognitive agents with rules for making decisions. Hierarchical structure is represented, most importantly from the modelling point of view, not by correspondence with the container hierarchy, but by the different kinds of interactions which take place between different kinds of agents. Similarly, agents are identified by the rules associated with their Agent subtype (e.g., Department Manager) as well as by their position in the container hierarchy.
This model has been devised to simulate the informal development of business processes within the organization. It is assumed that each process looks for a partner with which to combine in order to improve the value of their joint activity relative to the value of existing joint activities. The process which drives the dynamic processes is the search by individual agents for alternative activities with which they can combine to produce more value. Agents' search is not constrained by the formal departmental structure of the organization. One effect of formal organizational structure is to impose some order on the combination of activities. Activities undertaken within one department engage in some level of cooperation before cooperating with the activities in another department. But the individuals who undertake the activities might prefer to combine in a different order than that facilitated and supported by the formal structure and procedures.
To give an indication of the similarities and differences pertinent to an implementation of Soar agents in SDML, the paper both outlines the differences between them and indicates how the funcitionality of the Soar architecture can be implemented in SDML. To give the comparison concreteness, the Radar-Soar model (Ye and Carley, 1995) was implemented and run in SDML. However, the implementation was made as declarative as possible and, so, minimized the procedural element of cognition.
Ye and Carley report that the analysts and the manager each require some three hours of CPU time to complete an experiment with 60 aircraft where learning from feedback ceases after the first 30 are identified. The transcript from our experiments show that an experiment of the same length but with learning throughout takes less than 30 minutes including all decisions by all agents computed sequentially. Although our experiment was designed from the problem space map and flow charts reported by Ye and Carley, the much more declarative setup required one decision cycle for each aircraft and three elaboration cycles within each decision cycle. These cycles are not the same as in Soar since within each elaboration cycle the largest possible set of logically consistent steps are made. The SDML assumptions mechanism eliminates the inconsistent steps.
This assumptions mechanism was implemented both to make forward-chaining as efficient as (or possibly more efficient than) backward-chaining in (say) Prolog and to make SDML act within time slices as a formal logic. In particular, SDML corresponds to a fragment of strongly-grounded autoepistemic logic. Apparently, the forward-chaining mechanism of SDML (which also supports the correspondence to strongly-grounded autoepistemic logic) renders the whole process far more efficient than Ye and Carley achieved with Soar. Whether this is a property of Soar or the Ye-Carley setup or some more pervasive property of declarative and procedural modelling mixes is a matter for further research. A Soar-to-SDML compiler has been specified and will in due course be implemented in order to provide a framework within which to address that question.
A GARBAGE CAN SIMULATION MODEL OF HIGH-RISK FACILITY SITING
William E. Paterson
Department of Management
Carol Martin Gatton College of Business and Economics
Whereas past case studies have convoluted garbage can
with interpretive idiosyncracies, previous garbage can simulation models have explicated the details of the theory independently of case contexts. The siting process presents a set of interesting and difficult parameters (Carley, 1985) for a garbage can model. The parameters are interesting because they differ from the parameters in the Cohen, March, and Olsen (1972) garbage can model. The parameters are difficult because they are part of a real world phenomenon. Normative-rational decision making models that posit consequential relations between solutions and problems have been employed in the chemical demilitarization program by the U.S. Army for the siting of disposal facilities. In contrast, garbage can theory posits the decoupling of rational connections between solutions and problems through their simultaneity. Consequentially disengaged problems and solutions threaten the rationality of the siting process and may increase the likelihood of the reactivation of chemical weapons and munitions for military use.
A case context was used to parameterize and calibrate a garbage can simulation model of the siting process for a nerve agent incinerator proposed by the U.S. Army for the Lexington Bluegrass Army Depot in Richmond, Kentucky. Principal elements of the description and model of the siting process are the outcome termed the "siting decision" and the context of the siting decision. The siting decision consists of both a technology and its location. The destruction alternatives formally considered have included only incineration technology at different locations. In the garbage can model of siting, alternatives were reduced to on- and off-site destruction. In the simulation model, the siting decision outcomes, expressed as on- and off-site destruction alternatives, were linked with garbage can decision system performance measures. Additional parameterization entailed differentiation of choice opportunities and the operationalization of deadlines and drift. The specificity of the context-based garbage can simulation model suggests the robustness of a garbage can theoretical perspective to public choice decision processes. In contrast to Levitt and Nass (1989) where an institutional environment "constrained or put the lid on" (Levitt & Nass, 1989: 191) on a garbage can decision process, an anarchic process relaxed or took the lid off the institutional process of siting for public scrutiny.
A narrative and a garbage can simulation model were used to explore the effects of organized anarchies on the siting process. Two research hypotheses motivated the construction of the model. One problem was an account of problematic persistent alternatives during the siting process. A complementary problem was how feasible rival alternatives were introduced and competed as siting decisions. The classification of garbage can simulation models as discrete event models legitimated a network model of the siting process. Issue content (Padgett, 1980; Carley, 1985) internally differentiated the institutional and social-based choice opportunity energy supply and consumption patterns. Deadlines, considered in case studies (Weiner, 1976) as a moderating variable on the flow of garbage can stream variables, exacted time constraints on the behavior of solutions and problems in choices by undercutting the search for alternatives. Drift (March & Romelaer, 1976), as a garbage can system performance measure, represented the bias or predisposition of the context to alternatives with the ratio of the number of on-site decisions to the total number of on- and off-site decisions. The simulation model output data was analyzed with response surface regression to isolate decision contexts of persistent and rival alternatives and speculate about the future course of the siting process.
High-risk technology systems such as nuclear power plants, petrochemical plants, and air traffic control systems present facility sponsors, applicants, managers, and stakeholders with formidable problems of control once the installation is operating (Perrow, 1984, 1994). Perrow (1984, 1994) argued that organization theorists cannot design systems that can achieve such control. This control dilemma is central to the concerns of both those advocating and those contesting the proposed siting of the nerve agent incinerator. At the core of these concerns is the issue of the risks of operating these facilities. The siting decision partially depends on this management of risk and risk perceptions. Risk analysis is promoted as a primary tool for organizational decision makers, yet the subjective nature of risk analysis results in skepticism about its precise role in siting decisions. Perrow (1984, 1994) questioned the utility of risk assessments of high-risk facilities in the normal accident class of systems. Perrow (1984, 1994) argued that systems such as nuclear power plants, petrochemical plants, and air traffic control systems, because of their system characteristics rather than their toxic, explosive, or genetic dangers, are subject to eventual failures or normal accidents that result from tightly coupled components and complex (nonlinear) interactions among the system components. Coupling and the complexity of interactions are characteristics absent from the risk assessments of such technologies, yet major factors in the cause of failures.
The garbage can model of siting represents the reconciliation of the generalizability of garbage can theory (Hughes, 1985) and the detail it offers under conditions of space and time bounds (Bacharach, 1989). As part of his challenge for organization theory to link micro- and macro-levels of behavior in explaining macro-organizational phenomena, Perrow (1986) asserted the inapplicability of garbage can theory to large scale phenomena (Perrow, 1986: 139). This simulation model suggests otherwise. The idiosyncracy of garbage can processes was a problem of causal inference (Weick, 1985; Kosko, 1987; Padgett, 1980) for the garbage can theoretical explanation and model of the siting process. The suggestions offered by Bacharach (1989), Weick (1989), Eisenhardt (1989), and Osigweh (1989) for theory building and testing reinforce the calls for garbage can models to incorporate contextual features, constraints, and hybridized components. The causal inference problem created problems for falsifiability and utility (Bacharach, 1989); garbage can theory testing and building (Eisenhardt, 1989); and the construction of a plausible account (Weick, 1989) of siting that was both generalizable and context-specific (Osigweh, 1989).
The theme of a decision process with solutions temporally rather than consequentially attached to problems places garbage can models of decision making at the radical fringes of organization theory. March and Olsen (1985) pointed out that rationality permeates history, management, self-perceptions and myths. The idea of decisions disengaged from intentions threatens some interests who depend on the view that rationality underlies the behavior of individuals, organizations, and institutions. The discounting of rationality may be interpreted as an attack on the role of management in organizational affairs and, in turn, the stance of teachers, researchers, and writers who celebrate rather than critique (Alvesson & Willmott, 1992) the subject.
Understanding departures from rationality with garbage can theory in the siting of weapons and munitions disposal facilities may help to ensure that chemical demilitarization programs remain associated with peace-time rather than war-time activity. The identification of a taxonomy of organized anarchies in the siting of disposal facilities, as part of chemical demilitarization, may suggest intervention strategies for consequentially realigning solutions with problems and decision makers to minimize the likelihood that chemical weapons and munitions inventories will be remobilized. Examination of community resistance and compliance, as anarchic characteristics of siting, with garbage can theory, can contribute to the analysis of current failures in creating more democratic processes and informed public choice (Rood & Grimes, 1997). Future extensions of garbage can models are considered with the temporal and spatial properties of organized anarchies; different software paradigms for their expression; and the computer to simulate them.
MULTIPLE AGENT APPROACHES TO FRAUD DISCOVERY IN FINANCIAL STATEMENTS
Daniel E. O'Leary
University of Southern California
A multiple agent approach is developed to better understand the use of multiple agents in fraud detection, with a focus on financial statement fraud. Specialized agents are constructed to search financial information for particular characteristics and then make an initial evaluation as to the likelihood of fraud. Then a coordination agent and broker agents are investigated as vehicles for consolidating the individual agent findings and making an assessment as to the existence of fraud.
Why a Multiple Agent Approach to Fraud Discovery?
There are a number of reasons and settings for using multiple agents in fraud discovery. First, companies may have a number of different databases, e.g., in an internet or intranet setting. Different agents can be generated to interface with and gather information from those different databases. Second, different agents can be used to exploit off site server computing capabilities, in client server or internet/intranet settings. As a result, multiple agent approaches can be quite efficient. Third, when structured as a multiple agent problem, different agents can have very different knowledge structures so that specialized agents can be develope= d to solve unique problems. Particular groups can take control of generati= on of specific agents in order to provide appropriate expertise.
In any case, the agents have a limited view of the world of fraud discovery. However, additional capability to discover fraud is garned when the findings of two or more agents are compared and synthesized.
Coordinating Multiple Agents Used for Fraud Discovery
Although the agents can be designed to function semi independently they must be coordinated. In particular, the agent's activities need to be determined and the findings of the different agents need to be integrated into a cohesive understanding of whether it is likely that fraud has been generated.
Two basic models are explored. First, a broker model can take inputs fro= m two or more agents in an attempt to cull additional knowledge about fraud by integrating the results of the different agents. Second, a coordination meta agent can be made responsible for coordinating the findings from all the different agents in an effort to determine the existence of fraud.
Application Domain and Approach
The application of fraud discovery discussed in this paper is in the area of financial statement fraud, with specific interest in growth companies. In particular, individual agents have knowledge in a number of particular areas, including issues of
1. extent of growth or shrinkage of profits
2. extent of growth in financial information such as inventory and sales;
3. extent of increase or decrease in accounts where income manipulation can occur, e.g., sales return;
4. determination of consistency of policies used to account for research and development;
5. changes in liquidity through accounts such as long term debt, common stock and working capital;
6. qualification issues by the auditors.
Prototyping the Agents
Knowledge for multiple agents and the coordination of those agents is being captured for a prototyped system. First, a knowledge based is generated for each agent. For example, one agent is concerned with gathering information about the relationship between sales return and inventory and sales (#3 above). In addition, based on the conclusions that they can draw, each agent makes an assessment as to the likelihood of fraud in their specific knowledge area. Second, a knowledge base for a broker agent that takes inputs from two specialized agents is generated in order to provide additional fraud analysis capabilities. Third, a knowledge base for a coordinating agent designed to take input from each of the information gathering agents is created. This coordinating agent provides a model of agent interaction.
PIF
Yan Jin
ADDRESSES
Bernard Robert N. minusone@esg.colybrand.com
Robert N. Bernard
Coopers and Lybrand Consulting
Emergent Solutions Group
Room 3S-106
1301 Avenue of the Americas
New York, NY 10019-6013
212-259-1553
212-259-1337 (fax)
minusone@rocotillo.esq.colybrand.com
http://www.colybrand.com/esg
Bruderer Erhard brude001@maroon.tc.umn.edu
Prof. Erhard Bruderer
The University of Minnesota
Carlson School of Management
271 - 19th Avenue South
Minneapolis, MN 55455
612/625-0583
brude001@maroon.tc.umn.edu
Bruggeman Jeroen jeroen@gaston.ccsom.uva.nl
Professor Jeroen Bruggeman
CCSOM
Roeterseiland, Sarphatistraat 143
1018 GD Amsterdam
THE NETHERLANDS
31 205252800 (FAX)
jeroen@gaston.ccsom.uva.nl
Butts Carter ctb@andrew.cmu.edu
Carter Butts
Carnegie Mellon University
Department of Social and Dec. Sci.
Pittsburgh, PA 15213
(412)268-2869
(412)268-6938 (fax)
ctb@andrew.cmu.edu
Carley Kathleen kathleen.carley@centro.soar.cs.cmu.edu
Prof. Kathlen M. Carley
Carnegie Mellon University
Department of Social and Dec. Sci.
Pittsburgh, PA 15213
(412)268-3225
(412)268-6938 (fax)
kathleen.carley@centro.soar.cs.cmu.edu
Christensen Lars LCC@dnv.com
Lars Christensen
Department for Strategic Research,
Division for Technology and Products
Det Norske Veritas, HF8vik,
Norway
LCC@dnv.com
Christiansen Tore TCHR@dnv.com
Tore Christiansen
Department for Strategic Research,
Division for Technology and Products, Det Norske Veritas, HF8vik,
Norway
47-67-57-74-53
47-67-75-20 (fax)
Crowston Kevin crowston@syr.edu
Prof. Kevin Crowston
School of Information Studies
Syracuse University
4-206 Center for Science & Technology
Syracuse, NY 13244
315-443-16767
315-443-5806 (fax)
crowston@syr.edu
Edmonds Bruce b.edmonds@mmu.ac.uk
Bruce Edmonds
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building
Manchester M1 3GH
UNITED KINGDOM
44 (0)161 247 3886
44 (0)161 247 6802 fax
b.edmonds@mmu.ac.uk
Fridsma, Douglas B. M.D. fridsma@smi.stanford.edu
Douglas B. Fridsma, M.D
Section on Medical Informatics
Medical School Office Building, X-215
Stanford Univ School Of Medicine
Stanford, CA 94305-5479
(415)725-6699
(415)725-7944 (fax)
fridsma@smi.stanford.edu
Gaylard Helen h.gaylard@mmu.ac.uk
Helen Gaylard
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building
Manchester M1 3GH
UNITED KINGDOM
44 (0)161 247 3886
44 (0)161 247 6802 fax
h.gaylard@mmu.ac.uk
Guenther Oliver Guenther@parc.xerox.com
Oliver Guenther
Xerox Palo Alto Research Center
3333 Coyote Hill Road
Palo Alto, CA 94304
(415)812-4465
(415)812-4334 (fax)
guenther@parc.xerox.com
Hogg Tad Hogg@parc.xerox.com
Tad Hogg
Xerox Palo Alto Research Center
3333 Coyote Hill Road
Palo Alto, CA 94304
(415)812-4147
(415)812-4388 fax
Hogg@parc.xerox.com
Huberman Bernardo huberman@parc.xerox.com
Bernardo Huberman
Xerox Palo Alto Research Center
3333 Coyote Hill Road
Palo Alto, CA 94304
(415)812-4147
(415)812-4388 fax
huberman@parc.xerox.com
Jin Yan yjin@usc.edu
Yan Jin
University of Sourthern California
yjin@usc.edu
Kamps Jaap kamps@ccsom.uva.nl
Jaap Kamps
Applied Logic Laboratory,
Center for Computer Science in Organization and Management,
University of Amsterdam,
Sarphatistraat 143, 1018 GD Amsterdam, The Netherlands
31-20 5252538
31-20 5252800(fax)
kamps@ccsom.uva.nl
Klos Tomas B. t.b.klos@bdk.rug.nl
Tomas B. Klos
Faculty of Management and Organization
University of Groningen
P.O. Box 800, 9700 AV Groningen
The Netherlands
50-3637085
t.b.klos@bdk.rug.nl
http://www.bdk.rug.nl/mo/io/tomas
Kunz John C. kunz@cive.Stanford.edu
John C. Kunz
kunz@cive.stanford.edu
Kwon Yul K. yulkwon@leland.stanford.edu
Yul K. Kwon
Stanford University
P. O. Box 12023
Stanford, CA 94309
(415)497-6863
yulkwon@leland.Stanford.edu
Lee JuSung jusung+@andrew.cmu.edu
Ju-Sung Lee
Carnegie Mellon University
Department of Social and Dec. Sci.
Pittsburgh, PA 15213
(412)268-2869
(412)268-6938 (fax)
jusung+@andrew.cmu.edu
Levitt Raymond E. rel@cive.stanford.edu
Prof. Raymond E. Levitt
Stanford University
Department of Civil Engineering
Terman Engineering Center 298
Stanford, CA 94305-4020
(415)723-2677
(415)725-8662 fax
levitt@ce.stanford.edu
rel@cive.stanford.edu
Maiers Martin maiers@cs.umn.edu
Martin Maiers
The University of Minnesota
Department of Computer Science
200 Union Street S.E.
Minneapolis, MN 55455
(612)625-4002
maiers@cs.umn.edu
McKelvey Bill mckelvey@anderson.ucla.edu
Prof. Bill McKelvey
UCLA
Anderson School
110 Westwood Plaza
Los Angeles, CA 90095-1481
310-825-7796
310-206-3337
mckelvey@anderson.ucla.edu
Miller Sam
Stanford University
Department of Civil Engineering
Terman Engineering Center 298
Stanford, CA 94305-4020
Moss Scott s.moss@mmu.ac.uk
Scott Moss, Director
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building
Manchester M1 3GH
UNITED KINGDOM
44 (0)161 247 3886
44 (0)161 247 6802 fax
s.moss@mmu.ac.uk
Nasrallah Walid wade@cive2.stanford.edu
Walid Nasrallah
Stanford University,
Construction Engineering & Management Division,
Department of Civil Engineering
(415)723-1871
(415)725-6014 (fax)
wade@cive2.stanford.edu
Nooteboom Bart b.nooteboom@bdk.rug.nl
Bart Nooteboom
Faculty of Management and Organization
Groningen University
P. O. Box 800
9700 AV Groningen
THE NETHERLANDS
31-503633852
31-503633850 (fax)
b.nooteboom@bdk.rug.nl
O'Leary Daniel E. oleary@rcf.usc.edu
Prof.Dan O'Leary
University of Southern California
Finance and Management
School of Business
3660 Trousdale Parkway
Los Angeles, CA 90089-1421
(213)747-2815 fax
oleary@rcf.usc.edu
Pasquali Corrado mc8153@mclink.it
Corrado Pasquali
University of Genova
Via Piero Foscari 70
00139 Rome, ITALY
39-6-8125115
39-6-70452806 (fax)
mc8153@mclink.it
Paterson William E. (Bill) bad305@ukcc.uky.edu
William E. Paterson
Department of Management
Carol Martin Gatton College of Business and Economics
University of Kentucky
Lexington, KY 40506-0034
606-257-9064
606-257-9070 (fax)
bad305@ukcc.uky.edu
Peli Gabor g.peli@bdk.rug.nl
Gabor Peli
University of Groningen
Faculty of Management & Organization
P. O. Box 899
9700 av Groningen
THE NETHERLANDS
31-50-63-97-67
31-50-63-38-50 (FAX)
g.peli@bdk.rug.nl
Polos Laszlo laszlo@ccsom.uva.nl
Laszlo Polos
Applied Logic Lab
Universiteit van Amsterdam
Sarphatistraat 143
1018 GD Amsterdam
THE NETHERLANDS
31-20-525-2598
31-20-525-2800 (fax)
laszlo@ccsom.uva.nl
Sastry Anjali masastry@umich.edu
Prof. Anjali Sastry
University of Michigan Business School
710 Tappan Street
Ann Arbor, MI USA 48109-1234
(313)763-1591
(313)764-3146
masastry@umich.edu
Stein Roger M. steinr@moodys.com
Roger M. Stein
New York University and
Moody's Investors Service
99 Church Street
New York, NY 10010
(212)553-4928
(212)553-4805 (fax)
steinr@moodys.com
Thomsen Jan jthom@cive.stanford.edu
Jan Thomsen
Construction Engineering and Management
Terman Engineering Center 392
Stanford University
Stanford, CA 94305-4020
(415)723-1871
(415)725-6014 (fax)
jthom@leland.stanford.edu
Van Zandt Timothy tvz@princeton.edu
Prof. Timothy Van Zandt
Department of Economics
Princeton, NJ 08544-1021
Fax: (609) 258-6419
(609)258-4050
(609)258-6419 (fax)
tvz@princeton.edu
http://www.princeton.edu/~tvz
Vermeulen Ivar ivar@ccsom.uva.nl
Ivar Vermeulen
CCSOM
Roeterseiland, Sarphatistraat 143
1018 GD Amsterdam
THE NETHERLANDS
31 205252800 (FAX)
ivar@ccsom.uva.nl
Wallis Steve s.wallis@mmu.ac.uk
Steve Wallis
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building
Manchester M1 3GH
UNITED KINGDOM
44 (0)161 247 3886
44 (0)161 247 6802 fax
s.wallis@mmu.ac.uk