Discussion and Conclusion

Position #23 (FEN ``8/7p/8/p4p2/5K2/Bpk3P1/4P2P/8 w - -'') of the well-known test suite BT-2630 exemplifies what interior-node recognizers are capable of. This endgame position features little material on both sides (KBPPP vs. KPPPP) and it requires White to play ``g4'' immediately in order to win. The table below shows that with the help of its recognizers DARKTHOUGHT manages to solve the position after roughly a second in iteration #10 while running on a 500MHz DEC Alpha-21164a PC164 workstation. Without the recognizers, however, it takes DARKTHOUGHT 2:24 minutes, 47 million additional nodes, and 8 more plies of search depth to do so.

BT-2630: Pos. #23	With Recognizers	Without Recognizers
Solved in Iteration	10	18
Time to Solution	1 sec	144 sec
#Nodes to Solution	351,263	47,531,238
Search Speed	325K nps	327K nps

These impressive differences spring to a large part from the added knowledge of the recognizers. Any chess expert who knows about the great importance of knowledge in endgames will agree that this example is not artificially constructed. The recognizers of DARKTHOUGHT behave like this in countless other endgame positions with little material on both sides, too. But comparisons of different versions of the same program with and without recognizers do not allow for a sound quantification of the recognizer framework alone. They rather measure the combined merits of both the recognizer implementation and the added recognizer knowledge. Because this knowledge could also be implemented by other means than interior-node recognition (e.g. in the static evaluation function), we need to compare our recognizer implementation with full-fledged alternative implementations of the whole recognizer knowledge in order to measure the effectiveness of the recognizer framework per se.

We deem the usefulness of such comparisons far too low as to justify their substantial implementation efforts. The advantages of recognizers over other forms of the respective knowledge are quite obvious in our opinion. We are convinced that most readers will agree with us on this issue after reading the following brief summary about recognizers and their distinguishing features.

• Interior-node recognition applies the knowledge at the earliest possible time during the search. It can cause direct cutoffs of complete subtrees and it ``massages'' the game tree already on the way down of the search by tightening the alpha-beta bounds.

• Bounded recognizer results enable the painless implementation of inexact knowledge like ``this position is at least/at most a draw/loss/win for Black/White''.

• Recognizer functions encapsulate handy chunks of knowledge that are completely independent of each other.

• Accesses to endgame databases integrate perfectly with interior-node recognition (simply code them into appropriate recognizer functions).

• Successful recognitions reduce the number of expensive static evaluations by reusing the normally much cheaper recognizer scores instead.

• Efficient implementation schemes make our recognizer framework extremely fast such that it does not slow down the search at all.

• Material signatures are useful for many other important tasks beside the detection and selection of recognizers.

Last but not least, we invite all interested readers and especially the still skeptic ones to take a closer look at the latest test games of DARKTHOUGHT as published on our WWW pages at <http://wwwipd.ira.uka.de/Tichy/DarkThought/>. In many of these games DARKTHOUGHT easily outplayed its strong commercial opponents in endgames with little material on both sides. The portrayed strength of DARKTHOUGHT in such positions clearly springs from its high-speed endgame recognizers. They turned many objectively lost games into draws and correspondingly drawn games into wins. The strength of the recognizers also surfaced during several electronic discussions on the Internet. DARKTHOUGHT routinely solved difficult endgame positions much faster than most other strong chess programs involved in the testing.