Real-time Distributed Algorithms for Visual and Battlefield Reasoning

There has been substantial recent interest in integrating knowledgebased reasoning (KBR) and case-based reasoning (CBR) within a single system due tothe potential synergisms that could result. Here we describe our recent work investigatingthe feasibility of a combined KBR-CBR application-independent system for interpretingmulti-episode stories/narratives, illustrating it with an application in the domain ofinterpreting urban warfare stories. A genetic algorithm is used to derive weights forselection of the most relevant past cases. In this setting, we examine the relative value ofusing input features of a problem for case selection versus using features inferred viaKBR, versus both. We find that using both types of features is best (compared to humanselection), but that input features are most helpful and inferred features are of marginalvalue. This finding is surprising to us, but it supports the idea that KBR and CBR providecomplimentary rather than redundant information, and hence that their combination in asingle system is likely to be useful. INTRODUCTIONIn many application fields, expert-level problem solving naturally involvesreasoning from a combination of both general knowledge and individual past cases. Well-known examples of this occur in legal reasoning, medical diagnosis and management,military tactical planning, software engineering, and related areas. From the viewpoint ofthose developing AI systems intended as decision aids, the need for reasoning from bothgeneral knowledge and individual cases has led to substantial recent efforts to find waysto integrate these two approaches within a single framework (reviewed in Marling et al,2002), and this continues to be an active research area today. In this context, we are investigating the feasibility of creating an application-independent approach to interpreting multi-episode “stories” that combines a variety of AIreasoning methods (rule-based reasoning/deduction, cause-effect reasoning/abduction,Bayesian inference, constraint-satisfaction problem solving, etc.) with the retrieval of pastrelated cases. Our goal is to implement a system that, given an application-specificknowledge base plus a database of past cases represented in terms of the same features, isable to generate inferences about new situations by concurrently using both knowledge-based reasoning and examination of past cases. This is an ambitious goal that involvesaddressing a number of challenging issues related to understanding narration [Herman,2003]. A central idea in this is that specific human-readable knowledge descriptions,written in a simple but formal knowledge representation format, contain sufficient