Neural Fuzzy Agents for Profile Learning and Adaptive Object Matching

A neural fuzzy system can learn an agent profile of a user when it samples user question-answer data. A fuzzy system uses if-then rules to store and compress the agent’s knowledge of the user’s likes and dislikes. A neural system uses training data to form and tune the rules. The profile is a preference map or a bumpy utility surface defined over the space of search objects. Rules define fuzzy patches that cover the surface bumps as learning unfolds and as the fuzzy agent system gives a finer approximation of the profile. The agent system searches for preferred objects with the learned profile and with a new fuzzy measure of similarity. The appendix derives the supervised learning law that tunes this matching measure with fresh sample data. We test the fuzzyagent profile system on object spaces of flowers and sunsets and test the fuzzy agent matching system on an object space of sunset images. Rule explosion and data acquisition impose fundamental limits on the system designs. 1 Smart Agents: Profile Learning and Object Matching How can we teach an agent what we like and dislike? How can an agent search new databases on our behalf? These are core questions for both human agents and intelligent software agents. We explore these questions with the joint tools of fuzzy rule-based systems and neural learning. These tools exploit the filter and set-theoretic structure of agent search. An intelligent agent can act as a smart database filter (Grosky, 1994; Maes, 1994). The agent can search a database or a space of objects on behalf of its user. The agent can find and retrieve objects that the user likes. Or the agent can find and then ignore or delete objects that the user does not like. Or it can perform some mix of both. The agent acts as a filter because it maps a set of objects to one or more of its subsets. The agent is ‘‘smart’’ (Brooks, 1995; Maes, 1995a; Steels, 1995) to the degree that it can quickly and accurately learn the user’s tastes or object profile and to the degree that it can use that profile map to search for and to rank preferred objects. Figure 1 shows how an agent can learn and store user tastes as a bumpy preference surface defined over search objects. Agent search depends on set structure in a still deeper way. The search system itself may have many parts to its design and may perform many functions in many digital venues (Colombetti & Dorigo, 1994; Yamauchi & Beer, 1994). But at some abstract level the agent partitions the object space into two fuzzy Presence, Vol. 7, No. 6, December 1998, 617–637 r 1998 by the Massachusetts Institute of Technology Mitaim and Kosko 617 or multivalued sets with blurred borders. The agent partitions the space into the fuzzy set of objects that it assumes the users likes and into the complement fuzzy set of objects that it assumes the user does not like. All search objects belong to both of these fuzzy sets to some degree. Then the agent can rank some or all of the objects in the preferred set and can pick some of the extremal objects as its output set. The agent needs a profile of its user so that it can group objects and rank them. The agent must somehow learn what patterns of objects the user likes or dislikes and to what degree the user likes or dislikes them (Maes, 1995b; Rasmus, 1995). This profile is some form of the user’s implicit preference map. The user may state part of this map in ordinal terms: ‘‘I like these red flowers more than I like those blue flowers. I like the large purple flowers about the same as I like the small redwhite flowers.’’ The objects may be fuzzy patterns or fuzzy clusters in some feature space (Krishnapuram & Keller, 1993; Pal & Bezdek, 1995; Pal, Bezdek, & Hathaway, 1996). Microeconomic theory ensures that under certain technical conditions these complete ordinal rankings define a numerical utility function. The utility function is unique up to a linear transformation (Debreu, 1983; Hildenbrand & Kirman, 1976; Owen, 1995). So we can in theory replace the ordinal claim ‘‘I like object A at least as much as I like object B’’ with some cardinal relation u(A) $ u(B) and vice versa. The utility function u: O = R converts the ordinal preference structure into a numerical utility surface in an object space O of low or high dimension (Debreu, 1983; Hildenbrand & Kirman, 1976; Owen, 1995). The user likes the surface’s peak objects and dislikes its valley objects. We use neural fuzzy systems to learn the user’s profile or utility surface as a set of adaptive fuzzy if-then rules. The rules compress the profile into modular units. The rules grow the profile from a first set of sample data or question-answer queries and change the profile’s shape as the agent samples more preference data. The modular structure of the rules lets the user add or delete knowledge chunks or heuristics. Figure 1. Profile learning. A neural fuzzy agent learns a user’s utility surface as the user samples a database of classic paintings. The twelve bumps or extrema on the preference map show how much the user (or the agent that acts on the user’s behalf) likes or dislikes the 12 paintings. Here the evolving utility surface forms in the ‘‘mind’s eye’’ of a neural fuzzy agent based on nineteenth-century English philosopher John Stuart Mill. 618 PRESENCE: VOLUME 7, NUMBER 6