Insanity... or Tal style?

[Dave Gomboc]

I have been tuning eval parameters in Gaviota with "apparent" success. First ~800 fast games suggest a possible increase in ~70 elo points from the version it played in CCT11. However, look at this position to see what Gaviota played.

[d]r4rk1/1bqnbppp/pp1ppn2/8/2PNP3/2N1B1P1/PP3PBP/2RQR1K1 w - - 5 13

First move after the setup position and it played Nd5?!?!?!? What?

I do not know where the heck that came from, and clearly, the tuning did not work in this position or... did it?

Congratulations: your program is playing real chess. :-) If you were playing blitz chess with a friend, you could plunk down Nd5 in a second, say "Thematic!", punch the clock, and let your friend suffer while working out the consequences!

I'd be interested to know more about your evaluation tuning (both procedure and resulting parameter values).

Dave

[michiguel]

I have been tuning eval parameters in Gaviota with "apparent" success. First ~800 fast games suggest a possible increase in ~70 elo points from the version it played in CCT11. However, look at this position to see what Gaviota played.

[d]r4rk1/1bqnbppp/pp1ppn2/8/2PNP3/2N1B1P1/PP3PBP/2RQR1K1 w - - 5 13

First move after the setup position and it played Nd5?!?!?!? What?

I do not know where the heck that came from, and clearly, the tuning did not work in this position or... did it?

Congratulations: your program is playing real chess. If you were playing blitz chess with a friend, you could plunk down Nd5 in a second, say "Thematic!", punch the clock, and let your friend suffer while working out the consequences!

I'd be interested to know more about your evaluation tuning (both procedure and resulting parameter values).

Dave

I have ~6 million of relatively quiescent positions from comp-comp games. Of course, I know the outcome of those games and the result from the perspective of the "side to move" (0, 0.5, or 1). Just for this experiment, I modified the evaluation of gaviota to give me probability to win for the side to move (from 0 to 1). Then, I fit the eval parameters to minimize the difference between the probability predicted and the real result. This was the intermediate step to test other ideas. For instance, I intend to refit the parameters to the score that Gaviota will obtain after a short search with the parameters that I just got (and more iterations could follow).

This preliminary attempt sounds crude but apparently it seemed to work in this case (so far, after ~2000 fast games looks like Gaviota gained ~80 ELO points). This does not demonstrate that this system is great, but I guess it demonstrates how horrible my parameters were. Now that I pay attention to them, it is obvious and I should have figure out by myself if I had invested time to do a proper manual tuning. Still, with this method I got back some parameters that look fishy (70 cp for bishop pair seems too high). At any rate, I think that there is some potential in all this.

I tested in the past other things but they did not work. For instance, I tried to fit it to the evaluation given by stronger engines like Crafty and Yace (which have a command "score" and are stronger than Gaviota); but the parameters obtained were not good at all. Maybe the idea was flawed, maybe I did not have enough positions (~30 k). I don't know.

Miguel

[Dave Gomboc]

I have ~6 million of relatively quiescent positions from comp-comp games. Of course, I know the outcome of those games and the result from the perspective of the "side to move" (0, 0.5, or 1). Just for this experiment, I modified the evaluation of gaviota to give me probability to win for the side to move (from 0 to 1). Then, I fit the eval parameters to minimize the difference between the probability predicted and the real result. This was the intermediate step to test other ideas. For instance, I intend to refit the parameters to the score that Gaviota will obtain after a short search with the parameters that I just got (and more iterations could follow).

There's a bit of a reinforcement learning flavour to what you are doing. When you minimize the difference between the probability predicted and the real result, it may make sense to penalize early (but non-book) positions less than positions closer to the end of the game.

Still, with this method I got back some parameters that look fishy (70 cp for bishop pair seems too high).

I see what you mean, though they frequently can be worth half a pawn in the middlegame, and may be decisive in the endgame.

I tested in the past other things but they did not work. For instance, I tried to fit it to the evaluation given by stronger engines like Crafty and Yace (which have a command "score" and are stronger than Gaviota); but the parameters obtained were not good at all. Maybe the idea was flawed, maybe I did not have enough positions (~30 k). I don't know.

Perhaps your engine does not have a superset of the concepts of these engines, or your engine interprets the occurance of some features differently from these engines? Perhaps the 30k positions did not give balanced coverage of the domain space?

Dave

[fern]

No veo justficacion en esas movidas de gaviota. Tal no era suicida. Si gaviota logró de todos modos un juego decente es más bien por sus otras virtudes y a pesar de dichas movidas. Debe haber por ahí una línea de tu código que premia en exceso alguna variable posicional.

saludos
Fern

[Dr.Wael Deeb]

No veo justficacion en esas movidas de gaviota. Tal no era suicida. Si gaviota logró de todos modos un juego decente es más bien por sus otras virtudes y a pesar de dichas movidas. Debe haber por ahí una línea de tu código que premia en exceso alguna variable posicional.

saludos
Fern

I do agree here that Tal was no suicidal in his chess....
I also agree that this type of playing style that Gaviota is showing here is due to a special code line in the latest version of the program....
Interesting to say the least....
Dr.D

[fern]

¿Manejas el español o usastes algun software para traducir?
F

[michiguel]

No veo justficacion en esas movidas de gaviota. Tal no era suicida. Si gaviota logró de todos modos un juego decente es más bien por sus otras virtudes y a pesar de dichas movidas. Debe haber por ahí una línea de tu código que premia en exceso alguna variable posicional.

saludos
Fern

Por eso lo de "Insanity" en el titulo...

De todas maneras, Nd5 no parece ser un error garrafal, pero si una jugada tipica de "jugador de cafe", osado, audaz, y con unas copas de mas.

No hay una linea que justifique esta jugada. La valoracion individual de cada una de la piezas se suman para compensar la perdida de material. Obviamente, algo esta ligeramente exagerado (la calibracion de todo esto no es trivial).

Saludos,
Miguel

[Dr.Wael Deeb]

¿Manejas el español o usastes algun software para traducir?
F

Google Language Tools actualy
Dr.D

[michiguel]

¿Manejas el español o usastes algun software para traducir?
F

Google Language Tools actualy
Dr.D

Ok, this is the text I wrote. I edited to make sure that google translate it well. The interesting thing is that in the process, google forced me to write a better spanish

Ese es el motivo por el cual puse "Insanity" en el titulo...

De todas maneras, "Nd5" no es necesariamente un error terrible, pero sí un movimiento típico de un "jugador de club" que es osado, audaz, y con unas copas de más.

No hay ninguna línea en mi código fuente que justifique esta jugada. La valoracion individual de cada una de la piezas se suman para compensar la pérdida de material. Obviamente, algo está ligeramente exagerado (la calibración de todo esto no es trivial).

Saludos,
Miguel

[JChess]

Hello, very very interesting games. You are to be commended for your hard work on this project. Is there any possibility that you will make this latest version of your engine available for download on your web site?

Thank you for your time and keep up the great work!!