Some discussion in the rybka forum when MicroMax is involved

[Uri Blass]

See the following discussion from the rybka forum

http://rybkaforum.net/cgi-bin/rybkaforu ... id=3221#fp

direct link for the post that starts discussion about micromax from that discussion

http://rybkaforum.net/cgi-bin/rybkaforu ... 1#pid42858

Summery of my claims in that discussion is the following:

1)The difference in rating points between only material evaluation and micromax evaluation is clearly higher than the difference in rating points between Micromax's evaluation and rybka's evaluation

2)Big part of the reason that micromax has lower rating than rybka is not inferior evaluation but inferior search.

3)Using the search depth difference to evaluate difference between micromax and other programs is misleading because micromax see less tactics at fixed depth relative to most engines.

I believe that micromax evaluation with good search should be enough
to get CCRL rating of 2500.

Uri

[hgm]

Part of the weakness of uMax is that it is slow, because it cannot afford to use efficient programming techniques (like piece lists). But at large depths it is probably hurt most by having poor move ordering, driving up the apparent branching ratio. The IID does sort the best move in front (and at d=0 takes the MVA/LVA capture), but if this move turns out bad on deeper search (or is no good capture at d=0 because, say, the MVV/LVA move is QxR and the Rook is defended, while the PV move would be NxB), it is totally in the dark about what the second-best move might be, and just tries them in board-scan order.

In itself, the search algorithm is reasonably advanced. It has null-move pruning, and LMR which exempts hash move, captures an Pawn moves from reduction. It also has a check extension. But a major weakness is that it does not recognizes check in leaf nodes, and happily takes a stand-pat cutoff on eval score while it is in fact checkmated. This makes it lose one ply of depth in recognizing checkmates, which makes it vulnarable.

Eval-wise, the main problem is its bad King safety recognition: this is almost completely indirect by penalizing King moves that the opponent could force if the King is exposed. But that also means it does not recognize its King safety is demolished near the leaves.

The eval weakness that is most noticeable if you watch games is its total disregard for doubled Pawns. But strangely enough it hardly improves its results if I fix that.