Thematic move heuristic

[muxecoid]

Hello. After reading the introductory materials on chessprogramming wiki I got an idea that will probably not work. I hope someone can explain why it can't work.

In many chess textbooks by masters and grandmasters you can find the observation that chess players may spend long time analyzing a move which worked well in other games with similar patterns on board even if the move looks pointless to a freshman.

Suppose that we have a "middlegame database". For known piece to move, destination square and captured piece it will hold patterns that recognize positions where such move is "thematic".

Such knowledge could be used to avoid reduction of potentially interesting lines that have very low score at low depth. Maybe for move ordering.


Example: if there is a white pawn on e5 and knight on f3 master is likely to spend much time on analyzing Bxh7+ bishop sacrifice. If the pawn on e5 is black strong players will normally discard Bxh7 immediately.

The pattern can involve masks for pawn structure and piece positions, material on board, etc. for some cases even partial pattern match is enough to consider the move thematic.

[hgm]

Ideas like that usually don't work, because 99.99% of all positions searched by the engine have no 'reality value', in the sense that they don't even remotely look like anything that would ever occur in a Chess game played by beginners, let alone by GMs.

[FlavusSnow]

I think he means isolating specific patterns or pattern fields. For example, I'd guess more than 75% of my games have at least one of the player with three pawns in front of a castled king. An engine could detect this pattern and somehow reduce or extend specific moves.

But seeing as how computers are notably bad at pattern recognition, I would also guess that most algorithms would fail. I'd speculate that the nuances of the pattern would govern (e.g. three pawns in front of a castled king, but other piece development is lopsided in favor of the attacker - different in my mind than three pawns in front of a castled king and piece development is lopsided in favor of the defender).

[hgm]

But the point is that such patterns are not really important. In the majority of positions the only pattern that matters is hanging pieces. To have any impact you should work on the algorithm that decides which of the many hanging pieces you can typically take is best taken, and if it is really worth it, or whether you should prefer to save your own hanging pieces, or just leave them hanging. The positions where it matters how many Pawns there are in front of your King are too few in number.