cotix / abalone Goto Github PK

For valid testing of relative strength of two versions, you need to control search by time spent, not by search depth. A handfuld of games with 1 second per move is an ok indicator. This is the only way to truly know if a change was an improvement or not.

For example, compare the strength of a fast but weak evaluation function versus a slow but strong evaluation function. Or, when you compare forward pruning techniques that increase depth - big search depth is not always an indicator of strength.

If you add a stop criteria tested during search, you can utilize time allocated very precisely. You need to be careful to implement the stop test very efficiently.

When the stop-test is working, next step should be to add iterative deepening. This works very efficiently, if you have transposition tables (which you already have).

In the readme it says "performance is already up to par, or better then ABA-PRO".

Well, performance, what is that? To most people, better performance means stronger play (wins most of the time). I'm pretty sure this is not yet the case.

Performance and search depth are only related, if you don't loose quality of moves. Often it is a trade-off.

I suggest you remove that sentence.

For information, Aba-Pro is probably the second strongest Abalone program in the world. David Malek reported that his program, My Lovely Abalone (MLA) has beaten Aba-Pro in a tournament game (2007). I'm not aware of any official challenge, though.

MLA uses MTD(f) search, as far as I know.

The evaluation function is too weak, compared to variants of compactness.

I suggest you implement a (slow) eval func that computes compactness as closely as you can to Tino Werner's description. Basic compactness is strong (first used in 1990's), but Tino has a twist that makes it even stronger on middle and end game. Then try to beat that function with a fast evaluation function.

It is possible to compute compactness incrementially with very low overhead, on 61-byte arrays. I have not been able to find one as fast for bitboards. My best effort is 3 types of masks, counting bits along the 3 axes. This will reduce it to three 1D problems, which can be computed simultaneously.