Wrapping up

It's time to conclude this investigation and see what we've learned. The original question was "Can a spam filter play chess?". Clearly, the answer to this is yes, but making it play well is not so easy.

A crucial aspect I haven't touched on here are chess tournaments. The only way to reliably judge potential improvements in an engine is to make it play other engines with known strengths. Not all types of tournaments are appropriate for the dbacl chess engine - for example randomized initial positions and endgame puzzles are meaningless, because dbacl doesn't think ahead more than one half move, and it needs the full game history for matching patterns. Moreover, the fundamental behaviour is learned entirely from training sets. Thus dbacl's strength as a chess player is meaningless without reference to its training archive. However, if this archive is fixed, then incremental improvements to the algorithms can be evaluated in that context.

dbacl is able to learn some tactics simply by reading large collections of games. However, it seems that strategy is beyond its capabilities. Moreover, there are some fundamental limitations in treating a PGN format game like a text document: some information such as the values of exchanged pieces aren't easy to read off without keeping an imaginary board for replaying moves.

There are many ways to change the characteristics of the basic chess engine we've built here. Besides more complex capture heuristics, one can try to account for the length of the game (opening/middle/endgame), the difference in the number of pieces captured by each side, etc. Beyond that, the PGN gameline representation which we used here can be replaced with more informative symbol sequences. In principle, one could replace each move with a pictorial representation of the board such as FEN, but besides the added implementational complications, this causes difficulties because of dbacl's limit of 30 character tokens, and possibly statistical issues in recognizing similar but not identical board configurations.

Perhaps most interestingly overall, it should be remembered that dbacl doesn't think ahead like most chess engines do. Its successes and failures are almost entirely based on the historical record of the game as it develops and mimicry of training games, not at all on calculating moves and countermoves in the future.

Download the latest dbacl chess engine: dce.sh