Observations about Komodo 2.03

[Mike S.]

At first I want to say that Komodo is of course a great engine. Thanks for this strong freeware. I am using it as an analysis engine e.g. in Chess Hero, with the expectation that especially due to the participation of Larry Kaufman, it will give human master like evals combined with the typical computer power. It has not disappointed me.

Nevertheless, maybe I can add an observation which may help to improve Komodo even further. I don't know if it is an "immanent" problem (of pruning or whatever) or just coincidence, but it looks to me as if Komodo has a habit NOT to search a better move anymore, as soon as the current pm gets very good evals already. IF this is so, this may be good from a practical viewpoint for eng-eng games to get big ratings efficiently, but it is not something a user likes to see in analysis.

The following is from singlecore, P4/2.4 GHz 128 MB hash only, but I think the depths each are more than ok. The positions are not ultra difficult and I use them since more than 10 or 15 years.

Quick-02, Quicktest
[D]1q6/r4pbk/1r1p2pp/B2Pn3/Q2NP3/1p3P2/6PP/1R3RK1 b - - 0 1

Analysis by Komodo32 2.03 JA:

1...b2 2.f4 Rba6 3.fxe5 Rxa5 4.Qc2 Qb6 5.Qxb2 Qxb2 6.Rxb2 Bxe5 7.Rb4 Ra1 8.g3 Bxd4+ 9.Rxd4
-/+ (-0.77) Depth: 6 00:00:00 8kN
=/+ (-0.57) Depth: 9 00:00:00 79kN
1...Qa8 2.Nxb3 Rxb3 3.Qxb3 Rxa5 4.Qb7 Qa7+ 5.Qxa7 Rxa7 6.Rfd1 Nc4 7.Rb4 Nb2 8.Rb1 Ra2 9.Kf1 Bc3 10.Rb7 Kg7 11.Rc1 Na4 12.Rc7 Be5 13.R1c2 Rxc2 14.Rxc2 Bxh2
-/+ (-0.88) Depth: 9 00:00:00 96kN
-/+ (-1.04) Depth: 17 00:00:19 3980kN
1...b2 2.f4 Ng4 3.Nc6 Qa8 4.e5 dxe5 5.f5 Rxc6 6.Qxc6 Rxa5 7.fxg6+ fxg6 8.Rxb2 Ne3 9.Qxa8 Rxa8 10.Re1 Nxd5 11.Rb7 Nf4 12.g3 Nd3 13.Rf1 Nc5 14.Re7 Ra6 15.Kg2 h5 16.Rff7 Ne6 17.Ra7 Rc6 18.Ra4 Nd4 19.Ra2
-/+ (-1.19) Depth: 17 00:00:29 5979kN
-+ (-2.07) Depth: 21 00:03:04 38810kN

After some interactive analysis up to 3...Rb4:

Analysis by Komodo32 2.03 JA:

1...Qa8 2.Nxb3 Rxb3 3.Qxb3 Rxa5 4.Qb7 Qa7+ 5.Qxa7 Rxa7 6.Rfd1 Nc4 7.Rb4 Nb2 8.Rb1
-/+ (-0.93) Depth: 6 00:00:00 2kN
-/+ (-1.04) Depth: 9 00:00:00 9kN
1...b2 2.f4 Ng4 3.Nc6 Qa8 4.e5 dxe5 5.f5 Rxc6 6.dxc6 Ne3 7.fxg6+ fxg6 8.Rfe1 Rxa5 9.Qe4 Nd5 10.c7 Qa7+ 11.Kh1 Nxc7 12.Rxb2 Nd5
-/+ (-1.20) Depth: 9 00:00:00 13kN
-+ (-2.35) Depth: 14 00:00:02 332kN
1...Rxa5 {correct} 2.Qxa5 Nc4 3.Qc3 Rb4 4.Kh1 Na5 5.f4 Bxd4 6.Qd3 Qb5 7.Qd1 b2 8.f5 g5 9.Qh5 Qd7 10.h4 gxh4 11.Qxh4 Nb3 12.f6 Be5 13.Qe1 Rc4 14.Qe2 Rc1 15.Qd3 Nc5 16.Qf3 Rxf1+ 17.Rxf1
-+ (-2.48) Depth: 14 00:00:02 346kN
-+ (-2.94) Depth: 20 00:00:46 10092kN

Mike's Test 2.2,N, Mikes_22
[D]4k2r/r2n1p1p/pq3p2/4pN2/Q1B1p3/4P2P/PP4P1/3R2K1 w - - 0 1

Analysis by Komodo32 2.03 JA:

1.Bxf7+ Kxf7 2.Rxd7+ Kg6 3.Ne7+ Kg5 4.Nd5 Qb8 5.Rxa7 Qxa7
+/= (0.65) Depth: 6 00:00:00 8kN
1.Bb3 h5 2.Qxe4 Kf8 3.Qd5 Qe6 4.Qxe6 fxe6 5.Bxe6
+/- (0.74) Depth: 6 00:00:00 10kN
+/= (0.28) Depth: 6 00:00:00 17kN
1.Bxf7+ Kxf7 2.Rxd7+ Kg6 3.Ne7+ Kg5 4.Nd5 Qb8 5.Rxa7 Qxa7 6.Qxe4 Rb8 7.h4+
+/= (0.65) Depth: 6 00:00:00 17kN
+/= (0.55) Depth: 7 00:00:00 20kN
1.Bb3 Kd8 2.Qc4 Re8 3.Qxe4 Kc7 4.Rxd7+ Kxd7 5.Ba4+ Kc8 6.Bxe8 Kb8 7.Qd5 Rc7 8.Bxf7 Ka7 9.b3 Qb7 10.Qxb7+ Kxb7 11.Bc4 Rd7 12.b4 Rd2 13.Kh2 Rc2
+/- (0.81) Depth: 7 00:00:00 22kN
+- (2.66) Depth: 18 00:01:51 25556kN

After some interactive analysis just after 1.Rd6:

Analysis by Komodo32 2.03 JA:

1.Rd6 {correct} Qd8 2.Rc6 Ra8 3.Nd6+ Kf8 4.Nxf7 Qe8 5.Nxh8 Nc5 6.Rxf6+ Kg7 7.Qxe8 Rxe8 8.Rc6 Nd3 9.Nf7 Nxb2 10.Nd6 Rd8 11.Bxa6 Rd7 12.Be2 Ra7 13.Nf5+ Kf8 14.Bh5 Ra8 15.Rf6+ Kg8 16.Bf7+ Kh8
+- (5.07) Depth: 6 00:00:00 1kN
+- (5.04) Depth: 18 00:00:07 1605kN

I don't claim that Komodo's first choices are directly incorrect, but the engine fails to find the even better moves. Make three 2nd best moves in a row and you will lose against every top-10 engine...

This is not something which is simply solved by using a faster computers like i7, and/or SMP and so on, because there are also more difficult positions of that kind and then you'd have just the same problem on a higher difficulty level. I want my analyses to find the absolute best move (keeping in mind that this may not always be possible anyway, but as a goal to go for always), disregarding how big the advantage for the current move found already is. If it's #12 please don't stop searching for a #11. That's what analysis is about, certainly not about being satisfied with 3rd or 2nd best moves, disregarding if they are "already winning" or not.

[Mike S.]

I tried to construct a simple example where O-O-O is good. Komodo 2.03 IS capable of long castling, in principle:

[D]r2q1rk1/1pp2p1p/2npb1p1/p5P1/8/2PBB3/PPPQ1PP1/R3K2R w KQ - 0 13

Analysis by Komodo32 2.03 JA:

13.0-0-0 a4 14.a3 f6 15.Rh4 Ne5 16.Be2 Ra5 17.gxf6 Qxf6 18.Rf4 Qg7 19.Rxf8+ Qxf8 20.Qd4 Rd5 21.Qxa4 Rxd1+ 22.Bxd1 Nc4
+/- (0.92) Depth: 6 00:00:00 2kN
+/- (0.97) Depth: 15 00:00:17 2731kN

For comparison:

Analysis by Houdini 1.5a w32 T1:

(...)
13.0-0-0 f5 14.f4 Re8 15.a3 Qd7 16.Rh6 Bd5 17.Rdh1 Re7 18.c4 Bxg2 19.Qxg2 Rxe3 20.Qd5+ Qe6 21.Rxh7 Qxd5 22.cxd5
+/= (0.28) Depth: 10/21 00:00:00 81kN
+/= (0.54) Depth: 15/42 00:00:14 5272kN

Analysis by Shredder 11 SE:

(...)
13.a3 Ne5 14.0-0-0 Nxd3+ 15.Qxd3 f5
+/= (0.61) Depth: 7/18 00:00:00 98kN
+/- (0.71) Depth: 13/33 00:00:18 4216kN
13.0-0-0 Ne5 14.Be2 a4 15.a3 Bc4 16.f4 Bxe2 17.Qxe2 Nc6 18.Qg4 Qe7 19.Qh3 f5
+/- (0.80) Depth: 13/36 00:00:29 6969kN
+/- (0.82) Depth: 14/35 00:00:39 9304kN

[Mike S.]

Ok, your silence is very telling. I guess I have adressed substantial weaknesses of Komodo, and that is why this engine is limited in the struggle for #1. I accept it! Because being #2 or #3 is also a great achievement...

[tomgdrums]

Ok, your silence is very telling. I guess I have adressed substantial weaknesses of Komodo, and that is why this engine is limited in the struggle for #1. I accept it! Because being #2 or #3 is also a great achievement...

Hi Mike!

I think I read somewhere that Don and Larry are trying to make Komodo more useful for analysis of human games.

I have often found that when analyzing my own games (which are definitely at the patzer level ), that sometimes the "best" move as suggested by Rybka is often only slightly better than the move I found in post mortem. So an engine like Junior or Komodo (or Shredder) will sometimes help me understand that my game move was weak but my post mortem choice is better and winning even though not "best".

Now analyzing one's games at the master level is obviously a different story, but even then the degrees between best and 2nd best are often infinitesimal.

Of course in engine vs. engine best is best.

[fern]

Perhaps my experience and intutiion that Komodo, strong as it is, is not that sharp to kill me -he does it the same, but slower- as other do is asociated with what you say in your smart analysis.
Maybe a bit more of tactics could do the trick?

fern

[Dr.Wael Deeb]

Perhaps my experience and intutiion that Komodo, strong as it is, is not that sharp to kill me -he does it the same, but slower- as other do is asociated with what you say in your smart analysis.
Maybe a bit more of tactics could do the trick?

fern

I didn't have the chance to play Komodo yet as nowadays I play in the 2400 Elo section of my rating list,but....from the several games I've watched so far it seems to me that Komodo is most likely a solid positional squeezer rather than tactical....

Someday I'll play the beast and will post the games here....
Dr.D

[fern]

Docv, did you get Granmaster by Excalibur? It is an entertaining piece of equipment. Plays at near 2000 level, good enough to give fun and not good enougt to avoid defeats by a player as you, but still can give surprises. I have mauled him several times, BUt I have also been defeatd in some specially good days of the machine. There is the joy, my friend...


Fern, low level player, high level writer, medium level fornicator.

[Dr.Wael Deeb]

Docv, did you get Granmaster by Excalibur? It is an entertaining piece of equipment. Plays at near 2000 level, good enough to give fun and not good enougt to avoid defeats by a player as you, but still can give surprises. I have mauled him several times, BUt I have also been defeatd in some specially good days of the machine. There is the joy, my friend...


Fern, low level player, high level writer, medium level fornicator.

No I didn't purchase it yet due to severe lack of time lately but it's on my mind and will happen sooner or later

And yeah,these little monsters give a load of fun and that's what it's all about after all

Dr.D,a computer chess fighter with a lot of ass-kicking scars but still a fighter....

[lkaufman]

I think that you are just observing that Komodo, like all the top rated programs, does a lot of forward pruning, which especially means discarding or reducing sacrifices (or failures to win material) that don't look very promising. You should test other top engines like Stockfish, Rybka, Critter, and Houdini on these positions to see if the story is any different. Komodo should eventually find the best move with enough depth, as should these other programs. If we took out the forward pruning to make Komodo solve these problems the elo would drop hundreds of points and no one would even be talking about Komodo. Forward pruning is what made Rybka king for a long time, and now everyone realizes that this is the way to make a strong program.

[mclane]

i find it very funny that especially this radical forward pruning is the key to success. i remember how the FIRST dedicated chess computers HAD TO use forward pruning because of their SLOW hardware.

1981 Mark V had to use B-strategy to come deep into trees. although doing often tactical mistakes, it was enough in travemuende to win.

the old dedicated chess computers had to prune much because their speed was not high enough to
come deep enough into the tree without forward pruning.

Mephisto III by Nitsche and Henne can be seen as the master in this discipline.

then, with rising speed, the brute force machines took over.
this was when 6502 came to 4 mhz and more.

constellation 3.6, superconstellation, novag expert, mephisto B+P,
and the forward pruners were beaten.
and now...