Professional Documents
Culture Documents
The Role of Chess in Artificial Intelligence Resea
The Role of Chess in Artificial Intelligence Resea
net/publication/220814778
CITATIONS READS
15 1,179
5 authors, including:
David E. Wilkins
SRI International
77 PUBLICATIONS 2,223 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Robert Levinson on 31 May 2014.
Feng-hsiung H s u J o n a t h a n Schaeffer
I B M T.J. Watson Research Center Computing Science
PO Box 704 University of Alberta
York town Heights Edmonton
NY 10598 Canada T6G 2H1
T. Anthony Marsland D a v i d E* W i l k i n s
Computing Science SRI International EJ227
University of Alberta 333 Ravenswood Ave
Edmonton Menlo Park
Canada T6G 2H1 CA 94025
Abstract Panel S u m m a r y
The factors that make chess an excellent domain for AI
Our eminent researchers including John McCarthy, research include:
Allen Newell, Claude Shannon, Herb Simon, Ken
Thompson and Alan Turing put significant effort • Richness of the problem-solving domain.
into computer chess research. Now that comput- • Ability to monitor and record progress accurately
ers have reached the grandmaster level, and are through competition and rating, because of its well-
beginning to vie for the World Championship, the defined structure.
AI community should pause to evaluate the sig-
nificance of chess in the evolving objectives of A I , • Chess has been around for centuries - the basics are
evaluate the contributions made to date, and assess well-understood internationally, expertise is readily
what can be expected in the future. Despite the available and is (generally!) beyond proprietary or
general interest in chess amongst computer scien- nationalistic interests. Has been considered a "game
tists and the significant progress in the last twenty of intelligence." Many players of the game feel men-
years, there seems to be a Jack of appreciation for tally "stretched."
the field in the AI community. On one hand this is
• Detailed psychological studies of chess playing exist.
the fruit of success (brute force works, why study
These studies suggest that human players use differ-
anything else?), but also the result of a focus on
ent reasoning modes from those in current chess pro-
performance above all else in the chess community.
grams. Further, the reasoning modes are also used in
Also, chess has proved to be too challenging for
many other problem-solving domains.
many of the AI techniques that have been thrown
at it. We wish to promote chess as the fundamen- • Excellent test bed for uncertainty management
tal test bed recognized by our founding researchers schemes - the basis of most expert problem-solving.
and increase awareness of its contribution to date. The well-definedness and discreteness of the game
have led many to ignore this.
548 Panels
speed up the machine 1000-fold. The current computer Some fundamental AI questions that will remain are:
chess work will only underscore the need for better ways
• Given a patient and seemingly perfect teacher (that
of adding and manipulating knowledge reliably.
is a superior chess-playing machine), how should one
The defeat of the human World Chess Champion
use it to "teach" an Al-based learning program about
sooner rather than later w i l l help artificial intelligence.
strategies for playing chess (given that the rules of
This w i l l help to re-establish chess as an ideal problem
chess themselves are already perfectly known)?
domain for experimenting with the fundamental prob-
lems of artificial intelligence, as elaborated more fully • A related but perhaps simpler problem comes from
by Donskoy and Schaeffer [1989]. the realm of endgame play. Given a perfect N-piece
database holding an optimal move for each position
" E x p e r t I n p u t s " are Sometimes H a r m f u l (or perhaps only the length of the optimal sequence
Feng-hsiung Hsu from that position, or even less, whether the position
I B M T J . Watson Research Center was won), develop a program that can deduce a sound
set of rules or strategies for playing the endgame per-
Experience from the chess machine Deep Thought sug- fectly (or at lea#t better than any other expert).
gests that inputs from chess experts, while generally
useful, cannot be trusted completely. A good exam- • Given endgame positions which cannot be solved by
ple of this is Deep Thought's evaluation function. Sev- search or databases alone, deduce a plan or playing
eral changes by capable human chess experts failed to strategy that will transform the position into a known
produce significant improvements and occasionally even (win/draw) state. In one class of positions the re-
affected the machine's performance negatively. Here, maining pieces are held to few squares. Progress
human experts, along with their expertise, introduced can only be made by a freeing move that converts
their own prejudices into the program. One way of solv- a short-term loss of material (or perhaps position)
ing this problem is to l i m i t the type and the amount of to the achievement of a later, more significant goal,
expert inputs allowed into the program; in other words, Consider the Duchess-Chaos game [Frey, 1983, pp.
having an almost "knowledge-free" machine. The avail- 269-274], which is still thought to be beyond brute
ability of on-line high quality chess game databases force search. Related examples abound, for instance
makes this an attractive approach. Instead of having giving up a passed pawn on one side of the board to
the value of, say, an isolated pawn set by human experts win a pawn race on the other.
either explicitly or in functional form, one can simply • Given a well-defined threat (for example mate) deep
tell the program that isolated pawns are important fea- in the tree, identify un-examined moves at an earlier
tures and statistical procedures, with some additional level along the current path that have the potential to
expert inputs, can then be used to decide the functional deny the threat explicitly. This a form of dynamic re-
form and the proper weighting of the features in ques- ordering of moves, but is also (if no potential denials
tion. exist) a good forward pruning criteria - providing
That more responsibility for knowledge should be evidence to abandon this line of play.
placed on machines is consistent with recent efforts
to handle the knowledge acquisition problem in expert The first two projects rely on perfect domain knowledge
systems and also in memory-based reasoning schemes and the availability of an un-tiring teacher to whom
where knowledge is generated statistically rather than questions can be posed. The need for convergence to
a solution within some arbitrary or unreasonably short
relying on symbolic learning, abstraction or domain
time-frame will thus be eliminated. The learning mech-
modeling.
anisms used will have the benefit of drawing on results
obtained by exhaustive means. Even simple rote learn-
O p e n P r o b l e m s a n d Lessons f o r A I
ing, for example, has its place in Artificial Intelligence.
T . A n t h o n y Marsland Scherzer et at. [1990] have shown how to use the moves
University of Alberta made during a series of chess games to ensure that a
Based on predicted advances in computer technology, poor move sequence will eventually not be replayed.
particularly the faster speeds and increasing memory of By holding the computer's memory of played games
low cost systems, it is reasonable to assume that within in a hash table the information can be used to extend
the next decade the World Chess Champion will lose the depth of search during later play. In related work,
an informal game to a computer, and within twenty-five the expert knowledge from an encyclopedia of games
years lose a 12-game match. The early losses w i l l reflect has been optically read into a computer which also cor-
more breaks in concentration at first and later a recog- rected the typographical computer to play through all
nition of the inevitable, as arose when trains started to known games, identify "losing moves" and by back-
out-pace runners. Although the defeat of humans by tracking develop "innovations*' that correct the flaws. It
machines will be significant, it w i l l mean neither the is not trivial work because it requires making a plausible
destruction of chess as a pastime and learning medium, re-construction from imperfect data, but Ken Thomp-
nor the end of interest in computer chess per se. In- son has shown the way here [Marsland, 1987] Finally,
stead it will focus attention even more sharply on pre- once errors are found, a backtracking mechanism will be
cisely how and why humans can become so expert at needed to find the best place earlier in the game-tree
selecting sound (often optimal) variations in seemingly path to correct (avoid) the flaw that follows. The ap-
complex situations, without resorting to the exhaustive proach of recording, correcting and innovating appears
techniques used by computer programs. to be fundamental to Artificial Intelligence.
550 Panels
der, 1991]. Chess W a s G o o d for A I Research.
Morph stores two types of pattern: Graph patterns
David E. W i l k i n s
which represent attacks and defends relationships be-
SRI International
tween pieces and squares and Material patterns that
are vectors giving the relative material difference be- Over the years, chess has proven to be a fertile ground
tween the players, e.g. "up 2 pawns and down 1 rook," for ideas and techniques that have spread to other areas
"even material," and so on. Along with each pattern of A I . These include database enumeration techniques
is stored a weight that reflects the significance of the [Bratko, 1978], chunking [Campbell, 1988], search tech-
pattern. The weight is a real number in [0,1] that is niques (minimax, alpha-beta, iterative deepening), and
an estimate of the expected true minmax evaluation of the utility of information [Good, 1977], Considering
states that satisfy the pattern. the lack of funding for chess, it is significant that it has
produced so many results.
R e s u l t s w i t h M o r p h There have been many encour- Chess has been fertile because it provides a complex
aging signs in the three months since Morph was fully reasoning problem from a simple domain with a built-
implemented, and some preliminary results have been in performance criteria. The simple domain permits
published [Levinson and Snyder, 1991]. research to progress with little initial overhead. Having
R e l a t i o n s h i p o f M o r p h t o o t h e r approaches The a hostile opponent adds complexity to the reasoning. In
chess system combines threads of a variety of machine- many domains (natural language understanding comes
learning techniques that have been successful in other to mind), progress can be hindered by lack of perfor-
settings. It is this combination, and exactly what is mance criteria - it can be hard to tell whether the lat-
done to achieve i t , that is the basis for Morph's con- est thesis is an improvement on the current state of
tributions. The learning-method areas and their in- the art, Chess provides precise answers to performance
volvement in Morph include genetic algorithms [Gold- questions.
berg, 1989], neural nets (weight updating), temporal- However, hardware advances have made chess a less
difference learning, explanation-based generalization fertile ground for addressing the basic issues of A I . The
(EBG), and similarity-based learning. To combine game is small enough that brute-force search techniques
these methods some design constraints usually associ- have dominated competitive computer chess, and I see
ated with these methods are relaxed. W i t h genetic al- little AI interest in squeezing out the last few hundred
gorithms, structured patterns rather than bit strings points on the chess ratings, except for the psychological
are used. In contrast to neural networks the nodes impact of having a computer beat the human world
in Morph's hierarchy are assigned particular seman- champion.
tic/structural values. Temporal-difference learning is Obviously, many basic issues in AI are not naturally
usually applied to methods with fixed evaluation func- addressed in a game-playing environment and should
tion forms (in which the features are known but not be explored in other domains. These include commu-
their weights) but here the features change and the hi- nication, forming models of one's environment, sensor
erarchical database organisation produce atypical dis- analysis and integration, and (perhaps) reasoning about
continuities in the function. uncertainty. In addition, real-world domains force AI
Once a chess graph is constructed from a game board, researchers to address issues such as economy of scale,
the semantics of the nodes and edges in the graphs are noise, realtime response, failed actions, novel phenom-
unknown to the system. The only information a pat- ena, and multiple agents - issues that can be ignored
tern contains as far as the system is concerned is the in chess.
significance (weight) that has been attached to the pat- Of the AI areas well-suited to a game-playing do-
tern , in no place after pattern creation do we special main, there are better domains than chess. In partic-
case pieces or edges. Such a syntactic approach to the ular, Go has all the advantages of chess but provides
learning of search knowledge is substantially different more complex reasoning and an even simpler domain.
from many of the traditional symbolic AI approaches A successful symbolic Go program would have to plan,
to chess and to the learning of control knowledge. would have to use goal-directed search, would encourage
In summary, in addition to a unique combination of machine learning, and would promote visual reasoning
methods, what distinguishes Morph is: - all basic AI research issues that are now ignored in
competitive computer chess.
• A uniform representation of search knowledge.
Even better than Go, Barney Pell [1991] has proposed
• A syntactic approach to playing and learning. an event where programs compete against each other,
• An attempt to play a complete game of chess rather but are only given a description of the game to be played
than a small subdomain. at the beginning of the match. Chess is particularly well
suited to this adaptation. One could, for example, have
• Rejection of a learning-by-examples framework for
a competition using a chess board and pieces, where
an experiential framework that is more cognitively-
the match begins by giving the programs a declarative
inspired,
statement of now the pieces move, how they capture,
• Responsibility for feature discovery given to the sys- the initial position, and what the objective of the game
tem. is. The programs would have to play this newly de-
• Non-reliance on search (though at some point small fined game under time constraints. A longish series of
guided search may be incorporated, bringing us even games could be required. This would require machine
closer to the cognitive model). learning and a robust symbolic problem-solving capa-
552 Panels