Phase 1 Our prototype PARIS System relies on samples of information from three extensive bodies of previous, policy-oriented scholarship on the post-World War II UN system's efforts to manage and prevent violent inter-group conflicts - work by Lincoln Bloomfield, Ernst Haas, and Frank Sherman, plus their many associates. Specifically, it suggests further steps toward integrating Bloomfield-Moulton's pioneering computerized CASCON system for facilitating conflict management and prevention policy choices with the equally pioneering empirical scholarship of Ernst Haas on the past, present and future of the UN Collective Security regime. Haas inspired Butterworth, Alker and Sherman, additionally and with others, to develop more extended data sets, simulation models and analytical tools for trend analysis, historical retrospection and policy-relevant forecasting purposes. Sherman's SHERFACS data set, including Robert Butterworth's textual accounts originally inspired by Haas's work, is the most comprehensive of these by far. An upshot of this work was the recognition by Alker and Dwain Mefford that "case-based reasoning" tools could integrate these two lines of research in policy-relevant ways.

The present Web site is the principal result of the third phase of the ParisinLA project. The first phase was a proposal writing period, where the initial "white paper" proposal was written by Alker, with several emendations suggested by his USC Computer Science faculty colleague, Shankar Rajamoney. Alker and Rajamoney were co-principal investigators for the second phase of the ParisinLA project. Working partially on the basis of this White Paper, plus his earlier collaboration with Alker, and his own independent work, Mefford drafted a lengthier proposal for the Annenberg Center that suggested focussing on the arguments and analyses of conflict situations available from public media, especially those available in computational formats. Alker revised this draft at several points, principally to reinsert a focus on precoded informational bases, such as those he knew were available from Bloomfield, Haas and Sherman. Both Alker, Mefford and Rajamoney agreed that advanced textual analysis or "word modeling" computational techniques - such as the RELATUS system that Duffy and Mallery had developed at MIT and Hurwitz and Mallery's related computationally structured policy arguments - were beyond the time scale and resource limits of the ParisinLA project. The revised proposal was submitted to the Annenberg Center for Communication, and subsequently funded.

With the active involvement of Mefford, and the additional assistance of Rajamoney's graduate Computer Science students Jafar Adibi and Behnam Salemi, and Alker's graduate student Tom Vest, the project in its second phase, about 10 months, followed a two prong strategy overseen by Alker and Rajamoney. The first prong was the gathering, comparison and preliminary analysis of the CASCON, Haas, and SHERFACS data sets, with special concern with the conceptual and statistical properties of irregularly shaped, multi-phased, precoded case histories. The second prong of our analysis was the exploration of advanced, analytically integrated case-based reasoning modeling techniques, with an eye towards revising existing routines to fit a synthetic conceptualization of the inter-group conflict management/prevention domain.

 
Phase 2 Mefford's painstaking analytical reconstruction of CASCON's pre-publication case similarity measure led him to argue that its distance metric violated Euclidean axioms, and that it was too dependent on the way missing or "not relevant" factor codings were handled; this stimulated Al Mouton to revise CASCON's published format to allow analysts more weighting options, but it left ParisinLA participants looking for other approaches to distance/proximity assessment as well. Mefford's innovative demonstration of CASCON factor clusters in a non-metric multidimensional space emphasized their conceptual coherence, but also the heavy extent to which these factors were conceptually, analytically and empirically interrelated. Incentives were thus created for finding statistical or information-theoretic approaches to action-goal relevance assessment, eschewing causality arguments much favored by strict experimentalist or quasi-experimentalist research approaches.

With help from Salemi, Adibi undertook exploratory analyses of commercially oriented case-based reasoning systems, finding that typically they relied on inductive, nonlinear, criterion-based classification construction algorithms. Since Quinlan's reformulations of Hunt's earlier work on Concept Learning Systems, his ID3 and C4.5 classification learning programs, are publicly available and had been explored already by John Mallery and Frank Sherman, among others, it seemed highly appropriate to try applying appropriate variants of these inductive machine learning procedures to the CASCON, Haas and SHERFACS data sets. Adibi and Salemi's demonstration that C4.5 could be applied to ACCESS encodings of SHERFACS, combined with Mefford's chastening results on CASCON factor interdependencies, inspired Vest to develop a "register method" of summarizing the numbers of factors within CASCON's broad data categories leading towards or away from violence. These summaries of CASCON's judgmental codings were in turn shown using Quinlan's C4.5 to follow plausible classification rules suggesting which kinds of cases were regularly susceptible to UN/NGO management/mediation activities.

 

Given the positive suggestions of the utility of information-theoretically based, machine learning and "data mining" techniques, combined with the negative results of the more ambitious prong of our phase 2 research strategy, two months before the original project year was to conclude, Alker and Rajamoney, on the advise of Adibi and Mefford, decided that a Web site presentation of our work to date would be the best way to show what we had learned about practically oriented information support systems built up from, but not limited to, the texts and codifications of the CASCON, Haas-Butterworth and SHERFACS data sets. This was, in effect, largely a reversion to the emphases of the Alker-Rajomey White Paper preliminary proposal. A no-cost extension of the project was proposed and approved, with that understanding. The analytical tools we would try to illustrate, or incorporate into, our Web site were to include:

the ACCESS 97 relational data analysis package available in Microsoft Office 97 software, and necessary for exploratory cross-tabulations of coded variables;

the C4.5 classification-inducing package of procedures;

a set of flexible distance measures for quantifying case "proximities" and "distances" relevant to some target or outcome variable;

and some non-linguistic word matching procedures for judging overlaps in non-codified textual accounts, such as those briefly provided by Bloomfield and Moulton, or more extensively by Bloomfield and Leiss or Butterworth (and Sherman).

 
Phase 3 The last topic became the special concern of Mihaela Malita, a Fulbright visiting scholar at USC who joined the ParisinLA project team near the end of its second phase. She became an active member of the third phase of the project, along with Adibi, Alker and Vest. They were joined in the Spring and Summer of 1997 by Darold Higa, a graphics specialist, who helped significantly with graphics questions in our Web site design and in the retrieval of pictures to associate with Web site case descriptions. Malita, with help from Alker, spent several weeks as well in coordinating computer-readable versions of overlapping cases selected from CASCON, Haas and SHERFACS files. Adibi supervised the Web site development; Alker was primarily, responsible for the structure of the Web site, and the drafting most of its novel textual contents. Many of the specific decisions resulted from consensually-oriented working group discussions. An example of a late meeting of the ParisinLA working group is the following picture taken by Alker.In May, 1997.