aspiration windows

[Rentib]

I have introduced aspiration windows to my engine and surprisingly it resulted in about 30 elo loss. What's more, the results of the "go" command show different scores even without any fails (by fail I mean root score <= alpha or score >= beta). I reckon this is not what is supposed to happen and hence is my question: is this possible or have I just made some silly bug?

Here is my aspiration windows framework, which seems pretty normal in my oppinion:

Code: Select all

		alpha_window = beta_window = 25;
		while (true) {
			score = negamax(pos, ss, alpha, beta, depth);
			if (score <= alpha) {
				alpha = MAX(-CHECKMATE, alpha - alpha_window);
				alpha_window *= 2;
			} else if (score >= beta) {
				beta = MIN(CHECKMATE, beta + beta_window);
				beta_window *= 2;
			} else {
				break;
			}
		}
		alpha = MAX(-CHECKMATE, score - 25);
		beta = MIN(CHECKMATE, score + 25);

[alvinypeng]

Are you doing reduced window searches for every depth? Normally, you want the first few depths (maybe up to depth 5 so) to be with [-CHECKMATE, CHECKMATE] bounds. Then after that, use a reduced window.

[shawn]

What's more, the results of the "go" command show different scores even without any fails (by fail I mean root score <= alpha or score >= beta).

This is normal and expected since aspiration windows modify TT elements and interact with most search heuristics.

[Rentib]

This is normal and expected since aspiration windows modify TT elements and interact with most search heuristics.

Thanks a lot. I totally forgot that for example killer moves are changed due to alpha being higher.