Plumbing The Depths (after Round 44 of 54)

[Graham Banks]

PLUMBING THE DEPTHS

Intel Core 2 Duo E4400 @2GHz overclocked to 2.66GHz
2GB total RAM
Arena 1.1 GUI
HS-Mainbook.abk (v705)
256 MB hash for each
3-4-5 piece EGTBs
40/30 (adapted to CCRL 40/40)
6 cycles (54 rounds)
Run by Charles Smith


Standings after Round 44

29.5 - Adam 3.1
29.0 - OBender 2.11.5
28.0 - Gibbon 2.01b
28.0 - Buzz 0.07
25.0 - Feuerstein 0.4.51
21.5 - BigLion 2.23w
18.0 - Clueless 1.4
16.5 - Micro-Max 4.8
15.0 - Marvin 1.3.0
9.5 - Smash 1.0.3


The tournament can be followed and games downloaded from here:
http://kirr.homeunix.org/chess/discussi ... f=7&t=2595

[hgm]

micro-Max never ceases to amaze me!

Of course 16.5/44 is only 37.5%, meaning the opponents are ~90 Elo better.

But if you realize how little uMax knows about Chess, it is an absolute miracle that it can score against these opponents at all! (So far it was able to score against all, except Adam.)

The only Chess knowledge in there is:
- piece values,
- a general attraction of PNBK towards the center,
- hefty bonuses for Pawn on 6th and 7th,
- a Pawn-push bonus that slowly increases as piece-material is being burned,
- a discouragement to give up pawn control over a square,
- a penalty for moving the King if the board is still too crowded with pieces
- (except castling, which gets a bonus),
- and discouragement of moving away the 1-pawn pawn-shield directly in front of the King.

That is all! No open lines, no knowledge about passers, doubled or isolated pawns, no Rook-on-6th, no tailored piece-square tables. No end-game knowledge of any sort. No mobility. No real King-safety.

And still it can score against 2100-rated engines!

[Ovyron]

That is all! No open lines, no knowledge about passers, doubled or isolated pawns, no Rook-on-6th, no tailored piece-square tables. No end-game knowledge of any sort. No mobility. No real King-safety.

Are you planning on adding those things to Micro-Max to make some sort of Milli-Max?

[hgm]

I have no intension to incorporate features just for strength. The design criterion in the main development line will remain Elo/character. Dropping that, and going for pure strength, would make uMax just another engine, as this is what almost everyone is doing. And uMax just isn't a good starting point for that, almost anything in there you would do differently if there were no size limitations.

The funny thing is that the current evaluation seems quite well balanced, despite the things that are lacking. I have watched many fast games, and I cannot recognize an obvious shortcoming in its play. This in contrast to earlier versions, that had obvious blind spots, through which they were losing time after time. But after the last such hole I plugged, the strength did shoot up by an unexpected amount.

One thing that still seems an obvious error is its indifference towards doubled pawns. They almost always originate by recapturing something on 3rd rank with a pawn. Some versions ago I tried to improve this by discouraging pawn moves to squares that already had a similar pawn directly behind them. Strangely enough, that did not result in better gauntlet scores. But perhaps I should try again, as it is quite possible that smaller weaknesses are completely masked by bigger weaknesses. The matter of doubled pawns has always been somewhat puzzling to me, though. I recently discovered a bug in Joker, that had the sign wrong for the (sizable) penalty of doubled pawns for black, so that it was striving to create them (when playing black), rather than avoid them. Of course I immediately corrected it, but it did not seem to alter its Elo one bit. And Joker really does a thorough analysis of the pawn structure; this was not just a cheap trick to discourage their creation by recognizing a subset of the possible cases that could create them, and not caring at all how to get rid of them once you have them, as I tried in uMax.

The Rook-play already seems quite natural. uMax likes to occupy open files and 7th rank, despite the fact that there are no explicit evaluation terms to encourage it.

[Ovyron]

Dropping that, and going for pure strength, would make uMax just another engine, as this is what almost everyone is doing.

Yes, I knew that if you added anything to uMax, it wouldn't be uMax anymore.

And uMax just isn't a good starting point for that, almost anything in there you would do differently if there were no size limitations.

Have you considered starting again, doing things differently not limiting the size, or what are your plans for the future?

[Uri Blass]

Dropping that, and going for pure strength, would make uMax just another engine, as this is what almost everyone is doing.

Yes, I knew that if you added anything to uMax, it wouldn't be uMax anymore.

And uMax just isn't a good starting point for that, almost anything in there you would do differently if there were no size limitations.

Have you considered starting again, doing things differently not limiting the size, or what are your plans for the future?

He has another engine (Joker) when the target is playing strength.

Uri

[hgm]

Yes, I knew that if you added anything to uMax, it wouldn't be uMax anymore.

Well, I can add things, but they must increase the strength more than the size. It is not easy to find things that do. One thing I might still try is mobility in the evaluation. But if it doesn't add 100 Elo, it might not help, as I am not sure that I could do it with much less than 100 characters.