Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / what is the win rate in entirely random games?
- - By Chris Whittington (**) Date 2017-12-19 16:29
Does anybody know, or done the experiment?

How many games turn out to be wins (as opposed to draws) in a large sample of completely random move playout games. Program selecting entirely randomly randomly, game ending with mate or draw by FIDE rules (50 move rule, insufficient material, or mate).

A neural net can only learn if some of the random games give a win result.
Parent - - By Sesse (****) Date 2017-12-19 18:05
Did you consider Google? First page on “win rate random chess games”:

Seemingly the draw rate is only 84%.
Parent - - By Chris Whittington (**) Date 2017-12-19 18:20
it was slightly rhetorical ...
Parent - - By Sesse (****) Date 2017-12-19 18:25
If so, what was your rhetorical point? 16% win rate is more than enough to get a randomized neural network to start learning.
Parent - - By Chris Whittington (**) Date 2017-12-19 18:36
you just answered it. some people, I think they call themselves "suspicious", seem to be arguing that nothing can be learnt from blindly random-ing around games. But 1 in 6 or so games with a result will begin to get AZ neural net liking situations/lines where moves can be made adjacent to opponent king, and I think, without going back and looking at the paper again, that the move list, presumably side to move, is given as inputs. So, keeping to the self-imposed rule of the rules of FIDE chess, Deepmind have given AZ the kick start it might need. I wonder if it were just given piece positions without primitive, partial alternate move attack rays, whether it could have begun the unravelling process.
Parent - - By h.g.muller (****) Date 2017-12-25 18:52
The move list is not presented to the NN. Just the position, and a number of previous positions to recognize repetitions. The NN has outputs for every move that is ever possible in Chess. Ouputs corresponding to illegal moves are simply ignored.

As accidental mates are mostly inflicted by the side with the material advantage, and feedback will be evenly distributed over the entire game, the first thing the NN is likely to learn is that capturing and promoting is good, and then that LxH captures and Pawn pushing is good. And that strolling its King to the center is ill-advised. These are pretty simple patterns,  which almost certainly are already present in the initial random initialization of the weights, and then get quickly enhanced.

My non-searching engine N.E.G. plays by attacker / protector count and lowest attacker / victim value, plus a small preference for checking, and athat is enough to totally crush random movers.
Parent - By Chris Whittington (**) Date 2017-12-25 19:34 Edited 2017-12-25 19:42
yup, I misread the move list presentation as inputs. Realised that was wrong several days ago. But, nagging in the back of my mind is that I watched a video presentation in whcih I thought the speaker (Deepmind) said they did present moves. It's clear though from the paper that they didn't
Up Topic The Rybka Lounge / Computer Chess / what is the win rate in entirely random games?

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill