AI Magazine Volume 10 Number 2 (1989) (© AAAI)

CHESS REPORT Deep Thought Wins Fredkin Intermediate Prize

ince May 1988, Deep Thought almost all potential contenders. The S (DT), the creation of a team of two first place finishers drew with students at Carnegie Mellon Universi- each other. Hitech led the field at the ty, has been attracting a lot of notice. halfway point but lost to DT in round In the Fredkin Masters Open, May 3 and threw away a winning position 28–30, DT tied for second in a field of against Fidelity in round four (because over 20 masters and ahead of three of a programming bug). The level of other , including Hitech play in this tournament was by far the and Chiptest (the winner of the 1987 best ever in a event, and the North American Computer Champi- winners clearly deserved their top onships). In August at the U.S. Open, places. DT scored 8.5, 3.5 to tie for eigh- Ten days later, DT achieved the teenth place with greatest computer success to date. It among others. Its performance was tied for first with GM in marred by hardware and the prestigious Software Toolworks bugs. However, DT astounded every- Open in Los Angeles with a score of one by beating International Master 6.5, 1.5. Several GMs played in this (IM) Igor Ivanov, the perennial winner tournament, including former World of the U.S. Grand Prix circuit prize, Champion of the USSR. who is generally regarded to be as In the tournament, DT became the strong as the average first computer to beat a GM in a regu- (GM). lar tournament game when it beat In the American Open tournament GM , a former world title in Los Angeles in October, DT scored contender when was a modest 4.5, 1.5. However, three competing. It lost many times to U.S. weeks later at the U.S. Class Champi- Champion GM and onships in New Jersey, DT had an drew with IM McCambridge. DT impressive 5, 1, beating two IMs achieved a 6.5, 1.5 score against a (Bonin and Zlotnikov). Thus, it field with an average rating of 2492, a became the first program to beat more truly incredible performance. This than one IM. In the Hall of Fame tour- result moved DT’s U.S. Federa- nament in Canton, Ohio, in early tion (USCF) rating up to 2551, among November, DT won its first major the top 30 players in the United tournament, scoring 4.5, 0.5. During States. the tournament, it beat another IM After this event, DT qualified for (Blocker) and drew with co-winner IM the $10,000 Fredkin Intermediate Igor Ivanov, who was looking for Prize as the first computer to achieve revenge after what happened at the a USCF performance rating of 2500 U.S Open but was fortunate to escape over a set of 25 contiguous games in with a . human tournaments. The prize was A week later, in the strongest field awarded in January 1989 and was ever assembled for a shared equally among students Feng- event, DT tied for first with a special- Hsiung Hsu (head of the team and ly souped-up version of the top-of-the- designer of the hardware), Thomas line Fidelity Mach-III machine. The Anantharaman, Mike Browne, and winning score was 3.5, 0.5. This field Andreas Nowatzyk and Research 0738-4602/89/$3.50 © 1989 AAAI. also comprised Hitech, Cray Blitz, and Associate , a former

SUMMER 1989 00 student who is employed on the that operates within the framework of long as 30 ply are returned, and this Hitech project. DT’s record against the brute-force full-width alpha-beta kind of calculation can tax the ability GMs is now 1, 2 and against IMs 5, 2. search. This algorithm, called singular of the world’s best human players. DT These results mark its performance as extensions, makes it possible at mini- has announced a mate in 19 moves at least a strong IM. mal cost to detect that one move of a (37 ply) in an actual game, and DT is a dual-processor chess sibling set is clearly better than any Chiptest and Hitech have also per- machine in which 450,000 positions other alternative (Anantharaman, T., formed similarly, although somewhat per second can be done by each pro- Campbell, M., and Hsu, F. 1988. Sin- less deep. cessor and 700,000 per second by the gular Extensions: Adding Selectivity For those interested in more detail two working together. These figures to Brute Force Searching. Paper pre- on current computer chess activity, I make it the fastest chess machine sented at the 1988 AAAI Spring Sym- suggest joining the International ever, about five times faster than posium on Computer Game Playing, Computer Chess Association, which Hitech. DT’s predecessor, Chiptest, 22–24 March 1988, Stanford, Califor- produces a quarterly journal that cov- used only one of the move generator nia). Such moves play a powerful role ers these events in detail. Subscrip- chips in DT but was good enough to in determining forcing and forced tions (US $25/year) can be sent to win the 1987 North American Com- lines of play. Therefore, the algorithm , Department of puter Championships with a perfect does not count such singular moves as Computing Science, University of score in a field that included Cray a ply toward the prespecified depth, as Alberta, Edmonton, Alberta, Canada Blitz, , Lachex, and Phoenix. is normally done. The result is that on T6G 2H1. One reason (and we consider it the the average, an N ply search pene- major reason) for the recent excep- trates along the principal variation to Hans Berliner is a principal research scien- tional computer performances is the a depth of 1.5N and reaches a depth of tist in the Department of Computer Sci- discovery, by several members of the 3N about once in a game. This result ence, Carnegie Mellon University, Pitts- DT team, of a new search algorithm means that at times, variations as burgh, PA 15213.

00 AI MAGAZINE