Satana genetic learning: two days session results

[stegemma]

I'm running a long genetical tuning session, to tune the parameters of my engine. The results could be interesting, just to see how they change between generations and the dynamics of the resulting system. The full session can be downloaded from here:

http://www.linformatica.com/software/SatanaLearn010.zip

The engine play against itself with this parameters:

- 1000 generations limit (384 since now)
- 8 players
- 100 ms per move
- 100 half-moves limit

The score of any player is counted adding 1000 to the winner and subtracting moves count (faster wins are better). Loser lose the same value, of course.

The GA used is those described in previous posts with a change: any two player match starts from a random position, extracted from 1999-2000 FICS database; player a and b play from the same position, one time with white and then with black. This would give a more variation, even if the positions could be almost anything, from opening to final (even absurd ones!).

Tomorrow I'll try to apply the best set to Satana and see if it play better or not, against various version of itself and other similar level program (Neurone, Piranha, LarsenVB).

PS: the file is in ods format (Open Office calc)

[Daniel Anulliero]

Very good approach , interesting !
I think for a learning method too , for my engine Isa but not much ideas for the moment ...
"A suivre"

[stegemma]

Very good approach , interesting !
I think for a learning method too , for my engine Isa but not much ideas for the moment ...
"A suivre"

Let us know, if you find something new.

I've updated the file, now with the result of 617 generations. I'm testing the set of parameters against some other engine.

[Ferdy]

In every generation, how many games are played by each engine?

[Ferdy]

I've updated the file, now with the result of 617 generations.

Here is a plot to see overview after 618 generations. The Mass got big points at generation 161.

Code: Select all

#----------------- generation 161,,,,,,,,,,,,,,,,,,,,,,
161,1,Mass',46505,38,10,1,18,14,1,103,75,84,54,4,97,10,10,60,3,1,12,40
161,2,Eruption',9756,38,10,1,18,14,1,103,91,71,54,4,80,10,10,60,3,1,12,40
161,3,Iconoclast',6804,38,10,1,18,14,1,103,75,84,54,4,97,10,10,60,3,1,12,40
161,4,Stones of Years',6749,38,10,1,18,14,1,103,75,84,54,4,97,10,10,60,3,1,12,40
161,5,Battlefield',3971,38,10,1,18,14,1,103,75,84,54,4,97,10,10,60,3,1,12,40
161,6,Lucky Man',243,38,10,1,18,14,1,103,10,1,25,92,8,10,10,60,3,1,12,40
161,7,Manticore',35,38,84,1,18,14,1,103,75,84,54,4,97,10,10,60,3,1,94,40
161,8,Tarkus',-11987,38,10,101,25,28,75,34,57,70,9,30,15,24,82,100,67,1,12,40
,,,,,,,,,,,,,,,,,,,,,,

[stegemma]

Yes and looking at the graph it seems that the system oscillates in a pseudo-sinusoidal way. I think that 100 ms per move is a time too low for Satana, so the results are more random than expected. I could starts using a population the best ones to see what happens.

Your graph is very clean and it provide an instant sight of the dynamic of the system, thanks.

[stegemma]

In every generation, how many games are played by each engine?

One with white and one with black against any opponent (52 games in total). The FEN has been selected randomly for any single two games match (a vs b + b vs a with a single random FEN, c vs d, d vs c with another FEN).

[stegemma]

PS: maybe starting a session with the 8 best ones could be Embedded" in the algorithm itself; it would looks like a sort of Two-Layers Genetical Algorithm and we have the new chess programming acronym TLGA... but this is not the "Thailand Ladies Golf Association"