INTRODUCTION This paper is a report on some preliminary work undertaken as part of a longer term study of the problems which arise in designing and implementing digital computer programs which'learn'. A program has been written which learns to play the board game'Go-Moku' using a particular learning mechanism to be described later. The program is to be regarded as an experimental tool by means of which the particular learning mechanism can be investigated in some depth. Go-Moku is a simple but not a trivial game with an intellectual content comparable with a game of draughts (checkers). Opinions have sometimes been expressed that there is nothing to be learnt (no pun intended!) by programming simple games. Present knowledge of programming learning is such that it is useful to experiment with programs operating in a simple task environment. It is not so much what game the program learns as how it learns it. It is emphasised that the object of the present work is not to write a program which plays a difficult game better than anyone or anything has played it before, but to isolate and investigate particular aspects of a learning process which might be valid over a range of ill-structured problems. For the record, however, the current learning programs learn to play a good (basically defensive) game. The modifications currently being made to the program should give it a learning capacity to become unbeatable.

SEARCH TREES We will talk about problems with the following characteristics: (i) We could recognise a solution to the problem if given one. Let us call the set of objects which is known to contain the solutions the'candidates'. We include cases where there is more than one solution or where an optimal solution is required. Some examples are: (i) In playing chess there are only a finite number of possible strategies (candidates) but the number is far too large to enumerate. Any assignment is a candidate and there are a finite, but usually large, number of assignments. The candidates are the (n -- I)! permutations of the points omitting the starting point. In this section we will describe in an abstract way two approaches to problems of this type. We will give examples of their use, first in'Some problems about sets' (p. Although both the approaches have often been used before, the discussion may help to clarify those features common to the different applications. Our two approaches are both search techniques (partial enumeration techniques).

The author's main source was from personal discussion.)

RELATION TO PHYSIOLOGY It is possible to compare the organisation of this system with the organisation of the visual system, as revealed by microelectrode studies in the cat (Hubel & Wiesel 1962, 1965) and the frog (Lettvin, Maturana, McCulloch & Pitts 1959). Briefly the following points emerge: (1) Cells in the visual cortex only respond to local properties of the visual scene, e.g., edges, line segments. This mirrors the immediate constituent constraint imposed for economic reasons in the picture grammar.

Psychological theory attempts to explain how thinking--the subject matter of psychology--is possible by a brain composed of single mechanistic elements--the basic assumption of psychology. The problem of programming digital computers to behave in complex fashions is equivalent to this aspect of the psychological problem. Today automata theorists agree that no fundamental barrier blocks the development of machines which can think, by any reasonable definition of the term. However, the precise techniques for implementing general thinking proceGseE-J have been only partially developed. An example of a high-level, general thinking process is comprehension: the understanding of passages of a natural language.

The ability of the higher animals to accept and interpret information from distant objects confers enormous advantages for creatures (or machines) which respond only to immediate stimulation and have no opportunity to anticipate the future. Distance receptors, especially the eyes, serve as early warning systems by giving information of distance events, making it possible to gauge the probable future. The classical biological notion of stimulusresponse applies to creatures limited to touch information. The development of distance-receptors evidently allowed brains to develop to give strategic behaviour. It is unfortunate that the early, now classical, studies of reflexes involving touch and the internal regulation of the body have been so largely taken over to describe brain function, for these concepts are inadequate for describing the central nervous system. They tell nothing about how brains handle information from the eyes, to allow animals and man to see. They tell us nothing about decision-making: how present experience is related to the stored past to predict the immediate future.

Might computers one day be able to think every thought human brains can think? Might robots be able to behave in all the ways humans behave? Would such artifacts be fully human even though they had different structure and used different means? Since human evolution proceeds at a much slower pace than technological advance, might artifacts someday do everything better than people and become super-human beings? While computers and robots have made a promising start at overtaking humans at many thinking tasks and skilled behaviors, from chess to violin playing, these achievements have been made possible primarily by an exponential increase in computing power since the 1950s.

Using the Neomycin rule base, and inputting Neomycin's own actions to the action justification generator, the average size of J(.4,) was ten and the maximum size was approximately one hundred. When an Odysseus-generated rule base for the Neomycin domain was used, these set sizes increased by a factor of four to five. After the set J(Ai) is generated, the action justification ranking subsystem of Odysseus determines the likelihood that J(Ai) contains ji, the action justification of the specialist. This involves, first, ranking ji,„ in order of likelihood of being equal to the unknown An example of ranking rule is: given two elements of a J(.4,), where,4, occurs early in the problem solving session, the

LOGIC CROUP KNOWLEDGE SYSTEMS LABORATORY Department of Computer Science Stanford tIniversity Stanford, California 91305 Decision Precedures Abstract Distributed artificial intelligence is the study of how a group of individual intelligent agents can combine to solve a difficult global problem. This paper discusses in very general terms the problems of achieving this global goal by considering simpler, local subproblems; we drop the usual requirement that the agents working on the subproblems do not interact. We are led to a single assumption. An example of a distributed computation using these ideas is presented. Introduction The thrust of research in distributed artificial intelligence (DAI) is the investigation of the possibility of solving a difficult problem by presenting each of a variety of machines with simpler parts of it. The approach that has been taken has been to consider the problem of dividing tho original problem: what:,libtasks should be pursued at any given time? To which available machine should a..iven subtask be assigned? The question of how the individual machines should g9 about solving their subproblems has been left to the non-distributed Al community (or perhaps to a recursive application of DAI techniques). The assumption underlying this approach--that each of the agents involved in the solution of the subproblems can proceed independently of the others--has recently been called into question 12,3,6,7,101. It has been realized that, in a world of limited resources, it is inappropriate to dedicate a substantial fraction of those resources to each processor. The increasing attract:witless of parallel architectures in which processors share memory is an example of this: memory is a scarce resource. Automated factories must inevitably encounter similar difficulties. Are the robots working in such factories to be given distinct bins of component parts, and non-overlapping regions in which to work or to travel from one area of the factory to another?

Some of the more popular alternativo used to build knowledge systems are production systems, backward-chained reasoning, logic programming, heuristic search, and the Blackboard framework. Many of the applications implemented in production systems have been written in the OPS language [8]. In this framework, knowledge is represented as a set of homogeneous rules that are scanned for applicability in a data base that contains the current state of solution. Backward chaining also has a homogeneous set of rules, but the search for applicable rules is driven by a hierarchy of goals and sub-goals. The best known system for implementing this type of program is EMYCIN [4].