Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

An Answer

An Answer ICCA Journal 178 June 1994 CORRESPONDENCE The Editors received a question by Hermann Kaindl and an answer by Don Beal. Both are reproduced below in slightly edited form. A QUESTION by Hermann Kaindl Moebling, Austria With much interest I read the paper Random Evaluations in Chess by D.F. Beal and M.C. Smith in the ICCA Journal, Vol. 17, No.1, pp. 3-9. Having studied the problem of the pathology/benefits of minimaxing, I wondered whether the study reported in that paper may help to gain insight into this more general problem. Since the senior author was one of the fIrst to study this problem, I would be very interested in his opinion. by Don BeaP Queen Mary and Westfield College London, England Yes, I did consider the "random" experiment as contributing to the fundamental questions concerning minimax search, in the same spirit as pathology/benefits analyses. I believe the same effect will be observed in essentially all games that humans like to play. However, that statement has to be qualifIed by applying the analysis in slightly modified form for some games. For example, in Go the number of moves, equalling the number of empty intersections, tends to decrease steadily rather than fluctuate with the fortunes of each player, and is the same for both players (subject to minor effects from special rules). Since the "random lookahead" scores basically respond to the number of moves available, one might expect random lookahead to fail to produce benefits. However, in Go, a component of the final score is always available - namely the number of stones captured. It is arguable that this component should be included in any evaluation, even an otherwise random one. If it is, then the "random lookahead" benefits should be observable in Go. I have not done the experiment, though. Also, as I commented in the paper, I think the effect is significantly different from using number-of-moves directly as an evaluation, because "random" lookahead smears out the "measurement" of number-of-moves over all depths of the lookahead tree, rather than only measuring it at the horizon. I hypothesise that for very deep searches, random may do better than using "number-of-moves" at the horizon, despite the apparent greater efficiency of the latter. I also think the observation of random lookahead benefits is significant for completely unsupervised game learning. It means that learning systems can potentially start from absolutely nothing and progress, even in the beginning and middle phases of previously unknown games, far away from game-end results. 1 SchillerstraBe 45b, A-2340 Moebling, Austria. 2 Department of Computer Science, Queen Mary and Westfield College, Mile End Road, London El 4NS, England. Email: [email protected] http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png ICGA Journal IOS Press

An Answer

Beal, Don
ICGA Journal , Volume 17 (3): 1 – Sep 1, 1994
Download PDF
1 page

Loading next page...
 
/lp/ios-press/an-answer-INUhbQpCIW

References (7)

Publisher
IOS Press
Copyright
© 1994 – IOS Press and the authors. All rights reserved
ISSN
1389-6911
eISSN
2468-2438
DOI
10.3233/icg-1994-17319
Publisher site
See Article on Publisher Site

Abstract

ICCA Journal 178 June 1994 CORRESPONDENCE The Editors received a question by Hermann Kaindl and an answer by Don Beal. Both are reproduced below in slightly edited form. A QUESTION by Hermann Kaindl Moebling, Austria With much interest I read the paper Random Evaluations in Chess by D.F. Beal and M.C. Smith in the ICCA Journal, Vol. 17, No.1, pp. 3-9. Having studied the problem of the pathology/benefits of minimaxing, I wondered whether the study reported in that paper may help to gain insight into this more general problem. Since the senior author was one of the fIrst to study this problem, I would be very interested in his opinion. by Don BeaP Queen Mary and Westfield College London, England Yes, I did consider the "random" experiment as contributing to the fundamental questions concerning minimax search, in the same spirit as pathology/benefits analyses. I believe the same effect will be observed in essentially all games that humans like to play. However, that statement has to be qualifIed by applying the analysis in slightly modified form for some games. For example, in Go the number of moves, equalling the number of empty intersections, tends to decrease steadily rather than fluctuate with the fortunes of each player, and is the same for both players (subject to minor effects from special rules). Since the "random lookahead" scores basically respond to the number of moves available, one might expect random lookahead to fail to produce benefits. However, in Go, a component of the final score is always available - namely the number of stones captured. It is arguable that this component should be included in any evaluation, even an otherwise random one. If it is, then the "random lookahead" benefits should be observable in Go. I have not done the experiment, though. Also, as I commented in the paper, I think the effect is significantly different from using number-of-moves directly as an evaluation, because "random" lookahead smears out the "measurement" of number-of-moves over all depths of the lookahead tree, rather than only measuring it at the horizon. I hypothesise that for very deep searches, random may do better than using "number-of-moves" at the horizon, despite the apparent greater efficiency of the latter. I also think the observation of random lookahead benefits is significant for completely unsupervised game learning. It means that learning systems can potentially start from absolutely nothing and progress, even in the beginning and middle phases of previously unknown games, far away from game-end results. 1 SchillerstraBe 45b, A-2340 Moebling, Austria. 2 Department of Computer Science, Queen Mary and Westfield College, Mile End Road, London El 4NS, England. Email: [email protected]

Journal

ICGA Journal – IOS Press

Published: Sep 1, 1994

There are no references for this article.