Increasing attention has recently been paid to algorithms based on dynamic programming (DP) due to the suitability of DP for learning problemsinvolving control. In stochastic environments where the system being controlled is only incompletely known, however, a unifying theoretical account of these methods has been missing. In this paper we relate DPbased learning algorithms to the powerful techniquesof stochastic approximation via a new convergence theorem, enabling us to establish a class of convergent algorithms to which both TD("\) and Q-Iearning belong. 1 INTRODUCTION Learning to predict the future and to find an optimal way of controlling it are the basic goals of learning systems that interact with their environment. A variety of algorithms are currently being studied for the purposes of prediction and control in incompletely specified, stochastic environments. Here we consider learning algorithms definedin Markov environments. There are actions or controls (u) available for the learner that affect both the state transition probabilities, and the probability distributionfor the immediate, state dependent costs (Ci( u)) incurred by the learner.

Abstract: There exist a number of negative results ([J), [BR), [KV]) about learning on neural nets in Valiant's model [V) for probably approximately correctlearning ("PAClearning"). These negative results are based on an asymptotic analysis where one lets the number of nodes in the neural net go to infinit.y. Hence this analysis is less adequate forthe investigation of learning on a small fixed neural net.

Although the visual and auditory systems share the same basic tasks of informing an organism about its environment, most connectionist workon hearing to date has been devoted to the very different problem of speech recognition. VVe believe that the most fundamental task of the auditory system is the analysis of acoustic signals into components corresponding to individual sound sources, which Bregman has called auditory scene analysis. Computational and connectionist work on auditory scene analysis is reviewed, and the outline of a general model that includes these approaches is described.

Data clustering amounts to a combinatorial optimization problem to reduce thecomplexity of a data representation and to increase its precision. Central and pairwise data clustering are studied in the maximum entropy framework.For central clustering we derive a set of reestimation equations and a minimization procedure which yields an optimal number ofclusters, their centers and their cluster probabilities. A meanfield approximation for pairwise clustering is used to estimate assignment probabilities. A se1fconsistent solution to multidimensional scaling and pairwise clustering is derived which yields an optimal embedding and clustering of data points in a d-dimensional Euclidian space. 1 Introduction A central problem in information processing is the reduction of the data complexity with minimal loss in precision to discard noise and to reveal basic structure of data sets. Data clustering addresses this tradeoff by optimizing a cost function which preserves the original data as complete as possible and which simultaneously favors prototypes with minimal complexity (Linde et aI., 1980; Gray, 1984; Chou et aI., 1989; Rose et ai., 1990). We discuss anobjective function for the joint optimization of distortion errors and the complexity of a reduced data representation.

The 1993 International Logic Programming Symposium was held in Vancouver, British Columbia, on 26-29 October. It presented the state of the art in logic programming, emphasizing the deliberate interaction with other fields, in particular, humanistic fields. Topics covered at the symposium included algorithmic analysis, programming methodologies, semantic analysis, deductive databases, and programming language design.

MT Telecom, a Dutch telecommunications utility, has installed expert BNR Europe (Harlow, England), the instrument aboard the satellite. Pending system-based help desk systems to R&D subsidiary of telecommunications NASA approval, EUVE will be the centralize its 23 networked local data equipment supplier Northern first orbiting astrophysics mission to Telecom, is using virtual reality technology replace humans with AI technology. This installation proved to be a critical planning. The VR system allows Re:Member Data Services (Memphis, factor in helping the company BNR's engineers to visualize complex Tenn.), a data processor for obtain the IS0 9000 Total Quality installations and how they will work, credit union software services, has System Standard certification, a greatly saving time and effort compared automated all company service and requirement for those organizations to the traditional CAD system. Continental Bank (Chicago, Ill.) has expert system tracks all requests developed a client/server-based intelligent called in by users, and all requests Lockheed Missiles ST Space (Palo application to improve the can be accessed by anyone at the Alto, Calif.) has developed ASAP quality of its customer service.

Kambhampati presented theoretical planning systems is difficult. Although national AI conference, was lively It was noted that comparing planners encoding expert knowledge is at the and interesting. Both the theoretical is similar in difficulty to comparing heart of HTN planning, there and practical sides of the AI planning programming languages (in fact, the remains a considerable gap to bridge community were represented, input specifications to a planner can in using expert planning knowledge and both sides seemed to understand be viewed as a programming language). Shlomo Zilberstein (University Third, it was generally acknowledged Several papers contributed further of Massachusetts) presented a that common plan representations to the theoretical analysis of number of evaluation measures. A algorithms or through empirical An integrated system that executes common representation would allow studies (Christer Backstrom, or uses the generated plans formal comparisons among widely Linkoping University, Sweden; Subbarao should be evaluated instead of simply different planning technologies.

This paper is a multidisciplinary review of empirical, statistical learning from a graphical model perspective. Well-known examples of graphical models include Bayesian networks, directed graphs representing a Markov chain, and undirected networks representing a Markov field. These graphical models are extended to model data analysis and empirical learning using the notation of plates. Graphical operations for simplifying and manipulating a problem are provided including decomposition, differentiation, andthe manipulation of probability models from the exponential family. Two standard algorithm schemas for learning are reviewed in a graphical framework: Gibbs sampling and the expectation maximizationalgorithm. Using these operations and schemas, some popular algorithms can be synthesized from their graphical specification. This includes versions of linear regression, techniques for feed-forward networks, and learning Gaussian and discrete Bayesian networks from data. The paper concludes by sketching some implications for data analysis and summarizing how some popular algorithms fall within the framework presented. The main original contributions here are the decompositiontechniques and the demonstration that graphical models provide a framework for understanding and developing complex learning algorithms.

This report summarizes this year's workshop and outlines WET to underwrite and support these workshops. Information Systems is also acknowledged. The Defense Advanced this year's workshop and outlines the philosophy behind this annual event. Computer-Supported Cooperative and present the best research Finally, I would like to thank V. Work gathering, which takes in that has a bearing on the "repersonalization Jagannathan for his great help and everyone from anthropologists to of computing," as Fernando expertise in workshop management futurists, this workshop focuses on flores, founder of Action Technologies, and Mary Carriger for relieving me of hardware and software that enables puts it.

What criteria should be used to select one semantic formalism over another? However, the scope of analyze and gain insight into (that is, models for circumscription, perfect convergence results linking aspects of not just model) such a task. Although much basic problems are NP hard (at best). Ginsberg and Hugh Holbrook work remains to be done, the consensus His point was that just confirming (Stanford University) showed seems to be that there is sufficient that this problem is indeed potentially that default reasoning could be used common ground to warrant serious nasty is not really surprising. Marco Cadoli and as well as to somehow cope with the significant computational advantages.