Professional Documents
Culture Documents
The Future of Chess-Playing Technologies and The S
The Future of Chess-Playing Technologies and The S
net/publication/2826374
CITATIONS READS
12 717
1 author:
Dennis Decoste
eBay Research Labs
46 PUBLICATIONS 2,274 CITATIONS
SEE PROFILE
All content following this page was uploaded by Dennis Decoste on 29 August 2013.
Dennis DeCoste
Jet Propulsion Laboratory / California Institute of Technology
4800 Oak Grove Drive; Pasadena, CA 91009
http://www-aig. jpl.nasa.gov/home/decoste/, decoste~aig, j pl.nasa,
gov
10
not change as rapidly as Deep Blue’s might. and propagated up the tree such that the low estimate
Perhaps if Deep Blue wins an entire match against on the value of one movebecomeshigher than the high
Kasparov, it might become more acceptable to allow estimate of all alternatives.
such preparations, leading to more significant evalu- Since position evaluators generally only provide "re-
ation of in what ways the human mind may still be alistic" point estimates, B* uses null moves (GC90)
superior in the domain of chess. Thus, we argue that (and other heuristics) to generate optimistic and pes-
once (or if) a computer beats the world champion in simistic values. To avoid the complexity of managing
match, perhaps the real science will begin in earnest, full probabilistic distributions across these intervals,
although presumably with significantly less public in- they explored the use of probability distributions that
terest and drama. decay linearly from the realistic value to the optimistic
value. To handle instability of position evaluations,
Summary of Chess-Playing they use alpha-beta to conduct shallow probe searches
(e.g. full 3-ply, plus tactical extensions).
Technologies
Recently "rational search" methods have been pro-
In light of the diminishing returns of brute-force search posed for selective expansion based on Bayesian meth-
in chess noted above, we have considerable doubts on ods. This approach has been championed by Eric
whether the brute-force style epitomized by Deep Blue Baum (Bau92) (Bau93) (BS96). These methods
will actually suffice to prevail against Kasparov.In this tempt to identify at each iteration some small set of
section we highlight a variety of other AI techniques nodes whose expansion would be most informative,
that offer great promise in the longer run. based on detailed probabilistic distributions of the
evaluations of the current leaf nodes. Though gen-
Search eral and theoretically sound, use of these techniques is
incompatible with alpha-beta cutoffs. This approach
Perhaps due to the success of Deep Blue, many seem
to believe or to assume that parallel algorithms are the has not yet been demonstrated to be competitive with
prime future technology for computer chess. However, alpha-beta variants in the domain of chess, but cer-
tainly some reasonable approximations may prove use-
in fact, within the last decade there have been a wide
variety of advances in research on game-tree search. ful in the future.
It is important to keep in mind that, due to the de- Korf and Chickering has proposed an alternative
sign requirements that the Deep Blue hardware places selective search method called "best-first minimax
on its search algorithms, it is not clear which of these search" (KC94). Essentially, it always expands the leaf
techniques could be easily incorporated into the Deep of the current principal variation m i.e. the node which
Blue approach. determines the minmaxvalue of the root. They argue
Someelegant advances, though relatively modest that it will not generally suffer from excessive explo-
ration of a single path because of the well-known os-
performance-wise, have arisen from recent work by
cillation of values that occurs during minimaxsearch.
Platt et al. Their MTD(f) procedure (pSPdB95)
Best-first minimax search is relatively simple yet
(PSPdB96a) (Pla) essentially reformulates the
plex best-first method of SSS* (Sto79) into an elegant beats fixed-depth alpha-beta search in the game of
combination of zero-window alpha-beta searches and Othello for comparable-sized trees, for mediumdepth
extensive use of transposition (hash) tables. This work alpha-beta trees. However, deep fixed-depth alpha-
has led to modest yet surprising speed improvements beta search beats it for comparable-sized trees. This
for chess (15%) as well as significant clarity about suggests that its best use might be as a relatively shal-
low probe search for method such as B*, although Korf
alpha-beta and SSS* variants in general.
and Chickering did not discuss nor explore that option.
Their work on Enhanced Transposition Cutoffs
(ETC) (PSPdB96b) takes advantage of the fact Another promising area is that of generalizing the
notion of a transposition table so that each entry con-
game "trees" are actually graphs. For example, they
explore heuristics to maximize reuse of transposition tains an equivalence class of positions instead of just
table results, such as checking moveswith transposi- one. Ginsberg’s work on partition search (Gin96) has
developed this idea into the basis of a world-class
tion table entries first under certain conditions, since bridge program. Although he has not developed repre-
those might lead to immediate cutoffs. They report
sentations for chess, he notes that this might provide
that these heuristics lead to searches with 25%fewer
a more formal approach to the specialized though in-
nodes.
triguing "method of analogies" (AVAD75).
Less clear at this point, but potentially muchmore
powerful, are methods for selective tree growth. Re- Knowledge
cent results (BM96)on approximate probabilistic vari-
ants of Berliner’s well-known B* algorithm, for exam- Knowledge-engineering has long played a role in com-
ple, appear promising. B* involves search using in- petitive computer chess, particularly in the areas of
tervals to represent position values, instead of points. endgame table precompilation and manually crafted
Search continues until enough information is gathered opening books. Considerable effort typically goes into
11
selecting features and weights for the evaluation func- learning functions suitable for monitoring tasks were
tion. The requirement of such initial overhead to de- in fact inspired by ideas from B* search.
velop a competitive chess program has perhaps gone Webelieve and hope that such tight synergy between
a long way towards discouraging most AI researchers chess and "real world" domains might make serious re-
from even attempting, and towards computer chess’s search on chess more easily justifiable and even encour-
fall from grace. With the advent of mature search and aged. In the spirit of this workshop, we wish to stress
machine learning techniques, it seems likely that this that one of the significances of the GKvs DBmatches
will change in the near future -- particularly if the is the simple fact that there is renewed interested in
Deep Blue approach falls short. computer chess and thus it could becomea useful fo-
rum through which to educate the public about what
Learning AI is about as a whole. This gives added urgency to de-
velop more learning-intensive alternatives to the Deep
Recent advances in machine learning offer perhaps the
Blue approach, so that we are better able compare and
best hope for significant near-term advances in com-
contrast over spectra of approaches, rather than just
puter chess. Recent surveys such as (Fur96) and Jay
the human against machine theme per se.
Scott’s active web site at (Sco97) are good summaries
Detailed discussion of our techniques are beyond the
of a variety of approaches that have been proposed to
scope of this paper. However, we wish to outline our
date. Unfortunately, most of these approaches have
not been within the context of tuning a competitive- basic ideas here. In our monitoring work, we have
developed an approach called Envelope Learning and
level chess program, making their immediate impact
Monitoring using Error Relaxation (ELMER), which
difficult to assess.
essentially learns high and low limit functions for each
For example, there has been a fair amount of work
sensor, based on historic spacecraft data (DEC96).
on learning chess evaluation functions, with the Neu-
roChess work (Thr95) being perhaps one of the most Our techniques involve regression using asymmetric
interesting. cost functions (DeC97b),similarly to the 0-1 loss func-
tions commonin classification tasks. However,the key
Fktrthermore, there has been some work in learning
difference is that our cost function is parametric and
search control heuristics for game-playing, although
we search over those parameters and select the settings
the more successful ones seem particularly well-suited which give the best fit on cross-validation data.
for games with simplier features, such as Othello
In this way, we are able to learn envelopes which
(MM94).
It seems fair to say that machine learning to date do not necessarily suffer from classic problems such
has not had nearly the impact on computer chess that as variance underestimation (BQ97). In fact, our ap-
it could. It seems logical to expect that this situation proach is specifically designed to address cases which
will change as the impact of diminishing returns in are not handled well by the typical approach of learning
confidence intervals by first estimating means and then
chess search is better understood -- particularly when
research chess programs mature to the level that such estimating input-conditional variances (e.g. (NW95),
walls are routinely hit. (Bis95)). Thus, we can offer better assurance that
bound estimates are not too tight, as required both for
B* position evaluations and for spacecraft sensor limits
Summary of Our Chester Program for automated monitoring with few false alarms.
In this section we briefly summarizeour own research A simple example which illustrates the usefulness
on computer chess. Our search engine uses a variant of our approach is the task of learning high and
of B* to conduct high-level selective search, MTD(f) low bounds on the complexity of example run times
to conduct the relatively shallow low-level probes, and of an algorithm such as quicksort, given as input
extremely fast move generation and transposition ta- only the size N of each data set that was sorted.
ble routines, so as to reach a competitive level re- Consider that we wish to use linear regression (for
quired for meaningful evaluation. Wecall this program speed and for understandable results) and use the fol-
Chester (Chess Terminator), which reflects our serious lowing reasonable set of features for all such tasks:
but perhaps naive ambitions. The main novelty of our IgN, N, NlgN, N2, N3, 2N. With the high and low
approach is our means of attempting to learn good bounds of quicksort actually being (O(N2) and O(N lg
interval-valued evaluation functions, suitable for B*, N) respectively, it turns out to be difficult to capture
as opposed .to relying on heuristics such as null-moves these bounds using standard mean plus/minus stan-
to generate the bounds. dard deviation approaches. Essentially, the problem
The emphasis of our research is on using chess as is that there are critical missing inputs (namely, the
a domain to help test the generality of our machine "sortedness" of each data set to be sorted).
learning techniques being developed for real world Webelieve that such an approach is likely to prove
tasks such as monitoring the NASASpace Shuttle. useful for learning high and low bounds for chess posi-
(DeC97a)In reality, though, the causality of this line tion evaluation functions suitable for B* style selective
of research is not straightforward, since our ideas for search. However, we stress that we have not yet had
12
time to try to develop convincing results to support Andreas Junghanns, Jonathan Schaeffer, Mark Brocking-
this conjecture. Furthermore, we suspect that good ton, Yngvi Bjornsson, and Tony Marsland. Diminishing
context-sensitive evaluators are likely to require good returns for additional search in chess. In Advances in
feature construction techniques, such as the greedy in- Computer Chess VIII. (See http://www.cs.ualberta.ca/
troduction of products that we are currently experi- " jonathan/Papers/ dim.ps.Z).
menting with in our spacecraft domains (SSM92). Garry Kasparov. The day that i sensed a new kind of
intelligence. In TIMEMagazine. March 25, 1996.
References Richard E. Korf and David MaxwellChickering. Best-first
T.S. Anantharaman, M.S. Campbell, and F. Hsu. Singular minimax search: Othello results. In Proceedings of the
extensions: Adding selectivity to brute-force searching. Thirteenth National Conference on Artificial Intelligence,
Artificial lntelligence~ 43(1), 1990. pages 1365-1370, 1994.
G. Adelson-Velskiy, V. Arlazarov, and M. Donskoy. Some David E. Moriarty and Risto Miikkulalnen. Evolving neu-
methods of controlling the tree search in chess programs. ral networks to focus minimax search. In Proceedings of
Artificial Intelligence, 6, 1975. the Thirteenth National Conference on Artificial Intelli-
gence, pages 1371-1377, 1994.
Eric B. Baum. On optimal game tree propagation for
imperfect players. In Proceedings of the Eleventh National David A. Nix and Andreas S. Weigend. Learning local
Conference on Artificial Intelligence, pages 507-517, 1992. error bars for nonlinear regression. Advances in Neural
Information Processing Systems 7, 1995.
E. B. Banm. Howa bayesian approachs a game like chess.
In Games: Plannin 9 and Learning, AAAI Fall Sympo- Barney Pell. A strategic metagame player for general
sium, 1993. chesslike games. In Proceedings of the Thirteenth Na-
tional Conference on Artificial Intelligence, pages 1378-
Christopher M. Bishop. Neural Networks for Pattern
1385, 1994.
Recognition. Oxford University Press, 1995.
Aske Platt. Mtd(f), a new chess algorithm. WWW page:
Hans J. Berliner and Chris McConnell. B* prob-
http://theory.lcs.mit.edu/ " plaat/mtdf.html.
ability based search. Artificial Intelligence, 86(1),
1996. (See http:// www.cs.cmu.edu/ afs/cs.cmu.edu/ Aske Plaat, Jonathan Schaeffer, Wire Pijls, and Arie
user/ccm/www/home.html#B*). de Bruin. Best-first fixed-depth game-tree search in prac-
tice. In Proceedings of the Fourteenth International Joint
Christopher M. Bishop and Cazhaow S. Qazaz. Regression Conference on Artificial Intelligence, pages 273-281, 1995.
with input-dependent noise: A bayesian treatment. Ad- (See http://www.cs.ualberta.ca/ " aske/bfmms.ps).
vances in Neural Information Processing Systems 9, 1997.
Aske Plaat, Jonathan Schaeffer, WimPijls, and Arie
E.B. Bantu and W.D. Smith. A bayesian approach to de Bruin. Best-first fixed-depth minimaxalgorithms. Arti-
game playing. (See http:// www.neci.nj.nec.com/ home- ficial Intelligence, 1996. (See http://www.cs.ualberta.ca/
pages/eric/ bayesl.ps), 1996. " aske/aij-final.ps.gz).
Dennis DeCoste. Learning and monitoring with families of Aske Plaat, Jonathan Schaeffer, Wire Pijls, and Arie
high/low envelope functions: Maximizing predictive pre-
de Bruin. Exploiting graph properties of game trees. In
cision while minimizing false alarms. Technical Report D- Proceedings of the Fifthteen National Conference on Ar-
13418, Jet Propulsion Laboratory / California Institute of tificial Intelligence, pages 234-239, 1996. (See http://
Technology, March 1996. www.cs.ualberta.ca/ " aske/ AAAI96-final.ps.gz).
Dennis DeCoste. Automated learning and monitoring of Bart Selman, Rodney Brooks, Thomas Dean, Eric
limit functions. In Proceedings of the Fourth International Horvitz~ TomMitchell, and Nils Nilsson. Challenge prob-
Symposiumon Artificial Intelligence, Robotics, and Au. lems for artificial intelligence. In Proceedingsof the Fifth-
tomation for Space, Japan, July 1997.
teen National Conference on Artificial Intelligence, 1996.
Dennis DeCoste. Bounds estimation via regression with
Roger Schank. Where’s the AI? In AI magazine. Winter
asymmetric cost functions. Japan, August 1997. To be 1991.
presented at IJCAI-97 poster session.
Jay Scott. Machine learning in games. WWW page:
Mikhail V. Donskoy and Jonathan Schaeffer. Perspectives
http:// forum.swarthmore.edu/ " jay/ learn-game/ in-
on falling from grace. In Computers, Chess, and Cogni- dex.html, 1997.
tion. Springer-Verlag, 1990.
Terence D. Sanger, Richard S. Sutton, and Christopher J.
Johannes Furnkranz. Machine learning in computer Matheus. Iterative construction of sparse polynomial ap-
chess: The next generation. International Computer
proximations. Advances in Neural Information Processing
Chess Association Journal, 19(3), 1996. (See ftp:// Systems 4, 1992.
ftp.al.univie.ac.at/ papers/oefal-tr-96-11.ps.Z).
G.C. Stockman. A minimax algorithm better than alpha-
G. Goetsch and M. S. Campbell. Experiments with the beta? Artificial Intelligence, 12(2), 1979.
null-move heuristic. In Computers, Chess, and Cognition.
Springer-Verlag, 1990. Sebastian Thrun. Learning to play the game of chess.
In G. Tesanro, D. Touretzky, and T. Leen, editors, Ad-
Matthew L. Ginsberg. Partition search. In Proceedings vances in Neural Information Processing Systems, vol-
of the Thirteenth National Conference on Artificial Intel- ume 7, pages 1069-1076. The MIT Press, 1995.
ligence, pages 228-233, 1996.
Andreas S. Weigend and Neii A. Gershenfeld, editors.
F. Hsu, T.S. Anantharaman, M.S. Campbell, and Time Series Prediction: Forcasting the Future and Un-
A. Nowatzyk. Deep thought. In Computers, Chess, and derstanding the Past. Addison-Wesley, 1994.
Cognition. Springer-Verlag, 1990.
13