The importance of "Dynamism" in Komodo

[lkaufman]

I did some tests involving Dynamism. First, I ran huge numbers of three ply games with various settings to determine roughly what is optimum at that super-short level. I concluded that Dynamism = 75 and King Safety = 50 are about as good as I can do at three ply, and that these settings beat the default by forty elo points (at a 1.23% cost in time). These settings seem to give reasonable evals in positions where the default eval seems crazy. But when I ran these values against the default in timed games (15" +.15") the default won by some 66 elo, a net swing of 106 elo! So it seems that the optimum settings depend to a very great degree on search depth, although I don't yet know whether going out to more normal time controls would continue this trend or even reverse it. Anyway, it seems that these parameters do have major elo consequences. Any feedback on how alternate settings do at "real" time controls would be welcome. Maybe our users can help us improve Komodo.

[shrapnel]

I did some tests involving Dynamism. First, I ran huge numbers of three ply games with various settings to determine roughly what is optimum at that super-short level. I concluded that Dynamism = 75 and King Safety = 50 are about as good as I can do at three ply, and that these settings beat the default by forty elo points (at a 1.23% cost in time). These settings seem to give reasonable evals in positions where the default eval seems crazy. But when I ran these values against the default in timed games (15" +.15") the default won by some 66 elo, a net swing of 106 elo! So it seems that the optimum settings depend to a very great degree on search depth, although I don't yet know whether going out to more normal time controls would continue this trend or even reverse it. Anyway, it seems that these parameters do have major elo consequences. Any feedback on how alternate settings do at "real" time controls would be welcome. Maybe our users can help us improve Komodo.

Well, it does appear from the few Tests and games I have played, that the Komodo Team has indeed got the Default values quite correct for maximum ELO.
Further testing is required of course, but I have a feeling that any new values for these parameters won't be too far off from the already very strong Default values.

[lkaufman]

I did some tests involving Dynamism. First, I ran huge numbers of three ply games with various settings to determine roughly what is optimum at that super-short level. I concluded that Dynamism = 75 and King Safety = 50 are about as good as I can do at three ply, and that these settings beat the default by forty elo points (at a 1.23% cost in time). These settings seem to give reasonable evals in positions where the default eval seems crazy. But when I ran these values against the default in timed games (15" +.15") the default won by some 66 elo, a net swing of 106 elo! So it seems that the optimum settings depend to a very great degree on search depth, although I don't yet know whether going out to more normal time controls would continue this trend or even reverse it. Anyway, it seems that these parameters do have major elo consequences. Any feedback on how alternate settings do at "real" time controls would be welcome. Maybe our users can help us improve Komodo.

Well, it does appear from the few Tests and games I have played, that the Komodo Team has indeed got the Default values quite correct for maximum ELO.
Further testing is required of course, but I have a feeling that any new values for these parameters won't be too far off from the already very strong Default values.

Based on testing last night at 2' + 1", it appears that Dynamism = 110 might be a bit stronger than Dynamism = 100, which was best at much faster levels. This test used the latest dev. version of Komodo, but it is close enough to 9.42 for the result to be valid. It was plus 6 elo after 7000 games.

[JJJ]

I did some tests involving Dynamism. First, I ran huge numbers of three ply games with various settings to determine roughly what is optimum at that super-short level. I concluded that Dynamism = 75 and King Safety = 50 are about as good as I can do at three ply, and that these settings beat the default by forty elo points (at a 1.23% cost in time). These settings seem to give reasonable evals in positions where the default eval seems crazy. But when I ran these values against the default in timed games (15" +.15") the default won by some 66 elo, a net swing of 106 elo! So it seems that the optimum settings depend to a very great degree on search depth, although I don't yet know whether going out to more normal time controls would continue this trend or even reverse it. Anyway, it seems that these parameters do have major elo consequences. Any feedback on how alternate settings do at "real" time controls would be welcome. Maybe our users can help us improve Komodo.

Well, it does appear from the few Tests and games I have played, that the Komodo Team has indeed got the Default values quite correct for maximum ELO.
Further testing is required of course, but I have a feeling that any new values for these parameters won't be too far off from the already very strong Default values.

Based on testing last night at 2' + 1", it appears that Dynamism = 110 might be a bit stronger than Dynamism = 100, which was best at much faster levels. This test used the latest dev. version of Komodo, but it is close enough to 9.42 for the result to be valid. It was plus 6 elo after 7000 games.

Interesting for the next match that Graham will run I guess

[Jesse Gersenson]

In 5 seconds dynamism=190 finds tactical shots that dynamism=100 can not find in 5 seconds. But, generally, 190 plays worse.

Ok, so dynamism=190 is, generally, a patzer who shouldn't be trusted but who occasionally finds moves which change the game's outcome.

Getting a second opinion from the patzer sometimes pays off. Is there a way, or a set of conditions, which can get the patzer's opinion and not weaken the engine's play?

What are the scores of the tactical test suite for
a) 20 cores @ dyn.=100
vs
b) 1 core @ dyn=190

[Leto]

I was curious what does this mean:

The other new option, "Dynamism", gives the user full control over the style of Komodo; a low setting makes it prioritize static advantages like material and pawn structure, while a high setting prioritizes short-term ("dynamic") advantages like mobility and threats to the king. Although the default (100) is believed to be best for actual play, a lower value around 80 seems to produce more realistic evaluations in most positions so may be better for opening analysis.

I decided to use testsuites of different character to check for this setting. STS 1-15 - mostly positional in character, and Arasan 18 - mostly tactical.

STS: 5s/move

Code: Select all

Dynamism  Solved

 80        1346            
100        1367    ------
120        1341

The optimum is at the value of Dynamism 100.

Arasan: 5s/move

Code: Select all

Dynamism  Solved
100         126
140         155  
180         159
190         174    ------
200         171

The optimum is the value of Dynamism 190. The improvement compared to the default number of solved positions is huge, in fact we will see that combined with "King Safety", "Dynamism = 190" makes a tactical monster of Komodo.

So, in tactical problems, high values of Dynamism are favored. Maybe Larry's value of 80 is applicable to very quiet, positional problems?

Then, I also tested several hard tactical suites, and it is possible that Komodo 9.42 Dynamism=190; King Safety=160 (or similar values, try them) is the new King of tactics, maybe at least on a par with Houdini 4 Tactical, way above SF Matefinder. The testsuites were Arasan 18, HardTalkchess2015, DieRybka with about 5s/move.

Komodo D 190 KS 160 beats Houdini 4 Tactical in new Arasan 19 beta suite:
5s/move, 4 cores:
K: 178/250
H: 172/250

My impression is that by varying "Dynamism" from say 60 to 190, Komodo became both the strongest positional engine and the strongest tactical engine. Great feature!

I tried out your /d 190 KS 160 values in actual play (vs default) at 12" + .12", and after 272 games it is down by an astonishing 449 elo! I guess this shows that tactical problem sets have almost no correlation with playing strength.

This also brings up the question as to whether the perfect chess entity needs to be able to make the best move everytime. Perhaps it doesn't have to. Perhaps to increase likelyhood of winning one must focus on increasing positional advantage rather than seeking material advantage.

[shrapnel]

Based on testing last night at 2' + 1", it appears that Dynamism = 110 might be a bit stronger than Dynamism = 100, which was best at much faster levels. This test used the latest dev. version of Komodo, but it is close enough to 9.42 for the result to be valid. It was plus 6 elo after 7000 games.

Interesting. This seems to indicate that it is sufficient to fiddle around with the value for Dynamism, leaving King Safety untouched, to gain ELO, for a particular Time Control of interest.
That should save a considerable amount of time.

[Nordlandia]

Is dynamism of 110 recommended in engine matches?

I always tends to leave Komodo parameters at default, except table memory and contempt (adjust default 10 to 0)

[lkaufman]

Is dynamism of 110 recommended in engine matches?

I always tends to leave Komodo parameters at default, except table memory and contempt (adjust default 10 to 0)

We recommended the defaults at release time. However my present opinion is that you will get slightly better results with Dynamism = 110 at all levels except bullet levels; let's say 2' + 1" or more. This was confirmed by one of our testers at 30' + 15".

[Nordlandia]

Question related to Komodo's time usage heuristics.

Do ponder make experimenting with "Time Usage" parameter more appealing? assuming two computer face each others.

I was thinking about threshold range of -3 to +3 depening on advice.

Ponder might compensate for mild time usage adjustment, that was my idea.