Improving Source Selection in Analogical Reasoning An Interactionist Approach

The success of any analogical reasoner depends upon its ability to select a relevant source. We can improve source selection by more completely integrating the process of source retrieval with analogical inference, and by using experience in solving target problems to find properties that effectively predict a source’s relevance to future targets. This paper describes the design and evaluation of SCAVENGER, an analogical reasoning program that we have built to test these ideas. 1. Inleractk~snl and source retdeval Analogical inference is the process of reasoning about similarity: If two things are known to have certain similarities, we may infer by analogy that they are likely to have additional properties in common. Formally, an analogy is a mapping from elements of a source, a well understood problem solution, theory, plan, etc., to the target, a new problem to be solved. Analogical inference constructs this mapping, transferring knowledge from the source to the target in the process. The success of any analogy depends upon the selection of a relevant source. Most analogical reasoners share a common structure (Hall 1989; KedarCabelli 1988) that makes a number of assumptions about source selection. These assumptions include: 1. The separation of retrieval and inference. Many analogical reasoning programs treat source retrieval and inference as separate operations. They choose a source on the basis of known similarities to the target, prior to making any inferences. 2. Source oriented approaches to memory organization. To improve efficiency, analogical reasoners organize sources under a hierarchical index. Generally, they construct indices by comparing source descriptions to find properties that best distinguish them, and use these properties as indices for 179 source retrieval (Kolodner 1993; Fisher and others 1991). Our criticism of these assumptions derives from the interaction theory of metaphor (Black 1962). This theory views metaphors as complex interactions between systems of relations in the source and the target. These interactions, while primarily transferring information to the target, can also alter the semantics of the source. In particular, metaphors can create notions of similarity where none previously existed. Applying the interaction theory to analogical reasoning leads to the following criticisms of current source selection methods: 1. Source retrieval and inference should be combined. If metaphors and analogies can create similarities, how can we use similarity to select analogical sources? This circularity suggests that retrieval is itself a form of inference concerning the relevance of a source to a target. An analogical reasoning system should integrate retrieval and inference, choosing a source by evaluating the possible inferences afforded by alternative sources. 2. Index hierarchies should be constructed through experience in solving target problems. Most systems construct index hierarchies by comparing sources to find those properties that best distinguish them. Unfortunately, such algorithms can only consider a small number of properties in constructing hierarchies. The common approach to this problem is for the system designer to specify a fixed retrieval vocabulary. This ignores the reasoner’s experience in solving target problems. An alternative approach uses information gained in solving targets to find properties that will select relevant sources. 2. An Interactionist model of source r trieval We have formalized these ideas in a computer program, called SCAVENGER, that explains empiriFrom: AAAI Technical Report FS-94-02. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.