Analogy is like Cognition: Dynamic, Emergent, and Context-Sensitive

This paper presents several challenges to the models of analogy-making, namely the need for building integrated models, the need for using dynamic and emergent representations, the need for using dynamic and emergent computation, and the need to integrate analogy-making with other cognitive processes. Some experimental data are reviewed which substantiate these needs and the main ideas how the AMBR model of analogy-making could meet these challenges are presented. 1. From the Anatomy towards the Physiology of AnalogyMaking: The Need for Integrated and Dynamic Models For a long time now the research on analogy has concentrated on the anatomy of analogymaking, i.e. on decomposing it into pieces (representation building, retrieval, mapping, transfer, evaluation, learning) and trying to understand how each individual piece works. A number of successful models of various subprocesses (mainly of mapping and retrieval) have been built which account for most of the psychological data and make useful predictions: SME and MAC/FAC (Gentner, 1983, Falkenheiner, Forbus, Gentner, 1986, Forbus, Gentner, Law, 1995), ACME and ARCS (Holyoak, Thagard, 1989, Thagard, Holyoak, Nelson, Gochfeld, 1990, Holyoak, Thagard, 1995), IAM (Keane, Ledgeway, Duff, 1994), etc. The big challenge in modeling analogymaking (and human cognition in general) is to move on from the atomistic and analytical approach of Democritus (469-370 BC) towards the holistic and interactionist approach of Heraclitus (544-481 BC), i.e. to start building integrated models of the phenomenon as a whole. These models should unite contraries and account for data arising from the interaction between subprocesses, which cannot be explained by an isolated model of a subprocess. Such models are gradually emerging. Thus the CopyCat and TableTop models (Hofstadter, 1995, Mitchell, 1993, French, 1995) integrate representation building with mapping and transfer, LISA (Hummel and Holyoak, 1997) integrates access, mapping, transfer, and learning, AMBR (Kokinov, 1988, 1994c) integrates access, mapping, and transfer. Heraclitus took the view that “Everything flows, everything changes”, i.e. the dynamics of change is more important and informative than static objects and states. This is the next challenge to the current models: they should explain and predict not only the outcomes of the analogy-making process but also its dynamics. Unfortunately, only scare data is available on the dynamics of the process. This means that such data will have to be gathered by using experimental paradigms extensively used in other domains, for example, on-line experiments measuring reaction times, analysing thinking-aloud protocols, etc. These methods have already been used in analogy research but to a very limited extent (Ross and Sofka, 1986, Keane, Ledgeway, and Duff, 1994, Schunn and Dunbar, 1996). There are already experimental data which support the existence of interaction effects between the subprocesses of analogy-making. Thus Keane, Ledgeway, and Duff (1994) have demonstrated a very strong ordering effect, i.e. effect of the order of presentation of the target problem elements on the response time for solving the problem. Thus in the “singletonfirst” condition subjects found the mapping twice as fast as subjects in the “singleton-last” condition. These data can be considered as evidence for the interaction between perceptual and mapping processes. It would be even more interesting to find the reverse patterns: the mapping already established facilitating the perception of certain elements. The analysis of thinking-aloud protocols done by Ross and Sofka (1986) revealed that the retrieval of various elements of the source domain is interrelated with the mapping between the two domains, i.e. the already established mappings guide the retrieval of specific source elements. These data cannot be explained by a serial model of analogy-making where first the source is being retrieved and then the source and target are mapped. An extensive discussion of this phenomenon and its modeling in AMBR as well as simulation data obtained with AMBR can be found in (Petrov, Kokinov, this volume). AMBR predicts also the reverse influence: the specific order of retrieval of elements of the source domain will facilitate certain mappings. As a result of these interactions, a pattern of retrieval has been demonstrated where initially one source domain looks more promising and is better retrieved based on the greater superficial similarity, but as soon as mapping starts (in parallel to the continuing retrieval of domain elements), the higher structural correspondence between a second source domain and the target and the established mappings make it possible for the second domain to be ultimately better retrieved and mapped which would be impossible if the retrieval and mapping were sequential isolated and irreversible processes. Finally, a study currently underway involves video recording of subjects solving a formatting task on a computer screen. The video protocols demonstrate a complex interaction between perceiving elements on the screen (including figure/background perception), retrieving elements from memory, mapping between these elements, and performing actions on the screen, the results of which are further perceived and mapped to expectations. The explanation of all these data requires models which abandon the serial type of processing and which move on towards parallel processing which will allow the various subprocesses to interact dynamically with each other. AMBR is one such model that is based on the highly parallel cognitive architecture DUAL (Kokinov, 1994a, 1994b). All processes in AMBR are running in parallel and interacting with each other. Moreover, as described in section 3, each of these subprocesses emerges from the collective behavior of many micro-agents and thus is also inherently parallel. Since the micro-agents are taking part in various subprocesses there are no clear-cut boundaries between the various processes themselves. Before the dynamics of computation in AMBR can be presented, the need for dynamic representations that will change in the course of analogy-making will be discussed in the next section. 2. From Printed Text towards Moving Picture: The Need for Dynamic and Emergent Representations A printed text is a static representational object while a moving picture is a dynamic representation which emerges from the continuously changing frames. Moreover, this dynamic representation does not exist physically (only the static frames exist physically), it exists only in our consciousness. Analogously, memory traces may be considered either as physically existing static entities, or as emergent phenomena which are constructed in our consciousness. From the very beginning of memory research the view of memory as consisting of stable representations has been under fire. Thus Bartlett (1932) has shown that episodes are grouped into schemas and their representations are systematically shifted or changed in order to fit these schemas. Research on autobiographical memory has provided evidence that people modify their memories by dropping elements (schematising), including new elements (filling in), replacing elements (distorting), etc. Loftus (1977, 1979) has convincingly demonstrated a number of interference effects. One example involves subjects looking at a movie where a blue car does not stop at the site of an accident. Later on in a questionnaire a number of questions are asked about a different green car. As a result, when asked about the color of the car which did not stop, subjects are quite confident that it was green. In another study subjects claim they have seen broken glass in a car crash whereas there was no broken glass in the movie shown to them. Neisser and Harsch (1992) have demonstrated that the so-called “flash-bulb memory” does not exist but that descriptions constructed by human memory are so vivid that people strongly believe they are true. One day after the Challenger accident they asked subjects to tell them (and write down) how they learnt about the accident: whether they heard it on the radio, or saw it on TV, or learnt it on the street, in the supermarket, from friends. They asked further the subjects in the study what they were doing when they learnt about the accident, what their reactions were, etc. One year later the experimenters asked the same subjects whether they still remember the accident and how they learnt about it. People claimed they had very vivid (“flash-bulb”) memories about every single detail and they started to tell the experimenters a very different story from the one they told before. Even after the experimenters showed them their own writings they could not believe that the new story they were telling the experimenters was not true. Although it has long been demonstrated that human memory is a (re)constructive device rather than a store of stable memory traces from our past, models of analogy-making tend to ignore that fact. Typically these models would have a collection of representations of past episodes (prepared by the author of the model) “stored” in long-term memory (LTM), one or more of which would be “retrieved” during the problem solving process and would serve as a base (or source) for analogy. The very idea of having singular centralized and frozen representations of base episodes is at least questionable, but it underlies most analogy-making models, and certainly all casebased reasoning systems (Figure 1).