exact transposition table entries

[zenpawn]

My understanding is when an Exact type hash entry is found with adequate depth, it can be compared against beta to return a cutoff, and it can be compared with alpha to get a move and score. Shouldn't it also work to return alpha in the case that neither of those conditions is met? I tried this and it made my engine play much worse.

[Dann Corbit]

My understanding is when an Exact type hash entry is found with adequate depth, it can be compared against beta to return a cutoff, and it can be compared with alpha to get a move and score. Shouldn't it also work to return alpha in the case that neither of those conditions is met? I tried this and it made my engine play much worse.

If it hasn't met adequate depth, then you may be doing a zero ply search instead of what you requested. That unintentional pruning is rather savage, wouldn't you admit?

[Evert]

If you have an exact entry of sufficient depth, you should be able to return the score regardless of alpha and beta (modified for fail hard/fail soft).
Only in the case where the score is a bound do you need to compare against alpha and beta. If you do not have sufficient depth you do not return regardless of bounds (you need to search).

[zenpawn]

If it hasn't met adequate depth, then you may be doing a zero ply search instead of what you requested. That unintentional pruning is rather savage, wouldn't you admit?

I'm referring to the case where I have an Exact type entry with sufficient depth. When both score >= beta and score > alpha are not met, it would seem I should then be able to return alpha.

[zenpawn]

If you have an exact entry of sufficient depth, you should be able to return the score regardless of alpha and beta (modified for fail hard/fail soft).
Only in the case where the score is a bound do you need to compare against alpha and beta. If you do not have sufficient depth you do not return regardless of bounds (you need to search).

That's what I thought too. Since I use fail hard, I return beta if score >= beta, score if score > alpha, and to date have just fallen through and tried the move first otherwise. But, it occurred to me since it's an exact score of sufficient depth, I should know that it's already going to be worse than alpha and can save time by just returning alpha right there. Unfortunately, for some reason, this causes my engine to play much worse.

[bob]

My understanding is when an Exact type hash entry is found with adequate depth, it can be compared against beta to return a cutoff, and it can be compared with alpha to get a move and score. Shouldn't it also work to return alpha in the case that neither of those conditions is met? I tried this and it made my engine play much worse.

Too complicated. When you get an EXACT match, you return that score as if the current search had been executed and that was the final best score. The cutoff stuff should happen after you return. EXACT means "no searching or decisions need to be made here, just return the score and we are done."

[zenpawn]

Too complicated. When you get an EXACT match, you return that score as if the current search had been executed and that was the final best score. The cutoff stuff should happen after you return. EXACT means "no searching or decisions need to be made here, just return the score and we are done."

I'm using fail hard though, not fail soft.

Edit: Just tested this way and got the same results as my test of returning alpha should the other two conditions fail, so apparently it's equivalent. Not sure why it plays so much worse when the alpha case is not allowed to fall through.