Komodo knight odd match vs FM Victor Bolzoni : report

[Laskos]

We've tried a couple ideas out to capture the principle that positional factors seem more important as compensation for material than if they are just "free", but so far nothing like that has helped. Any ideas would be welcome.

Do you mean for example what I do in Andscacs, that it increases slighty the material imbalance bonus if there is big difference in mobility?

Code: Select all

//y = difference of mobility (from mg)
//z = compensation bonus, for mg and eg
			if (abs(y) > 50) 
				z = 15;
			if (abs(y) > 100)
				z = 25;
			if (abs(y) > 150)
				z = 45;

Now fires when there is material difference less than a knight. Needs to be improved and tested for bigger material differences, but already was a win.

Indeed Daniel, seems to work on Andscacs. On the other side, Komodo seems to need this sort of correction. I tested on 10 unbalanced openings, knight odds and positional advantage. These positions act as magnifying glass for the correction needed, in regular games the imbalances are much smaller. But I hoped for some at least 4-5 ELO points gain for Komodo even in normal games at slower than ultra-fast time controls. I maybe miss something.

Here are the positions:

Code: Select all

rnbqkbnr/pppppppp/8/8/2PPPP2/3BBN2/PP2Q1PP/R4RK1 w kq -
rnbqkbnr/pppppppp/8/8/2PPPP2/2NBB3/PP2Q1PP/R4RK1 w kq -
rnbqkbnr/pppppppp/8/8/2BPPB2/2N3P1/PPP1QP1P/1K1R3R w kq -
rnbqkbnr/pppppppp/8/8/2BPPB2/2N5/PPP1QPPP/1K1R3R w kq -
rnbqkbnr/pppppppp/8/8/2BPPB2/P4N2/1PP1QPPP/2KRR3 w kq -
rnbqkbnr/pppppppp/8/8/2BPPB2/5N2/PPP1QPPP/2KR3R w kq -
rnbqkbnr/pppppppp/8/8/2BPPB2/2N5/PPP1QPPP/2KR3R w kq -
rnbqkbnr/pppppppp/8/8/2PPPP2/3B1N2/PP4PP/R1BQ1RK1 w kq -
rnbqkbnr/pppppppp/8/8/2PPPP2/3B1N2/PP2Q1PP/R1B2RK1 w kq -
rnbqkbnr/pppppppp/8/8/2PPPP2/3BBN2/PP4PP/R2Q1RK1 w kq -

Komodo 9.1:

Code: Select all

1.5''+0.015''
Games        :    200 (finished)

White Wins   :     61 (30.5 %)
Black Wins   :     99 (49.5 %)
Draws        :     40 (20.0 %)
Unfinished   :      0

White Perf.  : 40.5 %
Black Perf.  : 59.5 %

6''+0.06''
Games        :    200 (finished)

White Wins   :     72 (36.0 %)
Black Wins   :     76 (38.0 %)
Draws        :     52 (26.0 %)
Unfinished   :      0

White Perf.  : 49.0 %
Black Perf.  : 51.0 %

24''+0.24''
Games        :    200 (finished)

White Wins   :     99 (49.5 %)
Black Wins   :     47 (23.5 %)
Draws        :     54 (27.0 %)
Unfinished   :      0

White Perf.  : 63.0 %
Black Perf.  : 37.0 %

Significant swing of Komodo of more than 150 ELO points to slower games.

Andscacs 0.82:

Code: Select all

1.5''+0.015''
Games        :    200 (finished)

White Wins   :     61 (30.5 %)
Black Wins   :    111 (55.5 %)
Draws        :     28 (14.0 %)
Unfinished   :      0

White Perf.  : 37.5 %
Black Perf.  : 62.5 %

6''+0.06''
Games        :    200 (finished)

White Wins   :     60 (30.0 %)
Black Wins   :    101 (50.5 %)
Draws        :     39 (19.5 %)
Unfinished   :      0

White Perf.  : 39.8 %
Black Perf.  : 60.2 %

24''+0.24''
Games        :    200 (finished)

White Wins   :     60 (30.0 %)
Black Wins   :     94 (47.0 %)
Draws        :     46 (23.0 %)
Unfinished   :      0

White Perf.  : 41.5 %
Black Perf.  : 58.5 %

Andscacs seems stable almost within error margins on these 10 positions.

Stockfish 17082015:

Code: Select all

1.5''+0.015''
Games        :    200 (finished)

White Wins   :     66 (33.0 %)
Black Wins   :    100 (50.0 %)
Draws        :     34 (17.0 %)
Unfinished   :      0

White Perf.  : 41.5 %
Black Perf.  : 58.5 %

6''+0.06''
Games        :    200 (finished)

White Wins   :     57 (28.5 %)
Black Wins   :     72 (36.0 %)
Draws        :     71 (35.5 %)
Unfinished   :      0

White Perf.  : 46.2 %
Black Perf.  : 53.8 %

24''+0.24''Games        :    200 (finished)

White Wins   :     60 (30.0 %)
Black Wins   :     53 (26.5 %)
Draws        :     87 (43.5 %)
Unfinished   :      0

White Perf.  : 51.8 %
Black Perf.  : 48.2 %

Swing of only 70 ELO points for Stockfish, which seems to need this correction less than Komodo.

[cdani]

Indeed Daniel, seems to work on Andscacs. On the other side, Komodo seems to need this sort of correction. I tested on 10 unbalanced openings, knight odds and positional advantage. These positions act as magnifying glass for the correction needed, in regular games the imbalances are much smaller. But I hoped for some at least 4-5 ELO points gain for Komodo even in normal games at slower than ultra-fast time controls. I maybe miss something.

Thanks for the test! Is really interesting and revealing.

Also I have in Andscacs a generic and increasing penalization for the side which is less developed.

At longer 20''+0.2'' the outcome already surpasses the eval of 0.00. My guess is that to longer TC it may be beneficial to even increase the importance of positional factors in Komodo. This might be a problem of all engines which have to pass first very short time control, it may harm there, but it probably scales well to longer TC.

Of course I do long tests on most positional parameter changes, even if the short tests shows as bad. There are more things than king safety and passed pawns that are good at long time controls. The outcome of the test of increase slighty the material imbalance bonus if there is big difference in mobility, also showed this pattern, better at long time control.

Also an engine tested at 60 seconds will be worst than one tested at for example 200 seconds, if they use the same test system and number of games to test, because it will detect some things that can only be good at very long time controls.

This explains also part of the chess engine advance in latest years, as the simple fact that now we use more powerful computers to test is equivalent to use more time of test in older computers.