Monte Carlo Data

[chrisw]

Chess also has some exit conditions apart from Mate and Stalemate:
1. Draw due to 3-move repetition
2. Draw due to 50 move rule

I guess Rybka Monte Carlo analysis would be using these exit clauses to end the games quicker.

But, there may be some games that may go very long, in case all options/moves are tried.

Yes, the problem is 49 moves then move a pawn, loop. And if the eval is not-knowing (as would be the case if thus were a learning technique) itr gets worse.

Contrast to, say, Backgammon, where the game progresses to conclusion naturally, even with random moves.

[Anil]

For unsing Monte Carlo analysis in chess, the engine eval would be very useful to keep the games to a short length. Search/Eval Depth of engine also plays an important role.

Like everything, the engines have been evolving and are continuously improving their eval/search.

[chrisw]

For unsing Monte Carlo analysis in chess, the engine eval would be very useful to keep the games to a short length. Search/Eval Depth of engine also plays an important role.

Like everything, the engines have been evolving and are continuously improving their eval/search.

Yes. Chess has a goal-driven termination and the evaluation function is designed to have knowledge of that goal and get there.

By contrast Backgammon is a one-way street and terminates naturally.

So the learning technology works for backgammon, a random eval function still get to the end of the game. Not so chess, Monte Carlo is learnable from for chess, but I doubt one can start from a no-knowledge base. Or maybe you can, if that never ending game poroblem can be dealt with?

Of course the other reason it works for backgammon is the randomising influence of the dice. Look-ahead in backgammon is not so useful, a statistical evaluation based on large numbers of games is the backgammon evaluation function via neural net technology.

Has Rybka managed to use statistical evaluation as its actual evaluation? That would be good and account for much strength imo.

[bob]

It was useful to reveal that white has no winning chances, at least.

"no winning chances at a very fast search pace". Many programs will fail to find fine#70 if you limit them to 15 plies or less, and they will probably draw every last time because Kb2 is much more natural than Kb1.

[bob]

I have mentioned this previously. In Cray Blitz, we tried _lots_ of ways of trying to automate the book selection process. We started by typing in all the moves from MCO 10 and MCO 11. Then we added BCO. Now we had a ton of opening lines, but with bogus analysis in places and even typos that turned a winning line into a losing one. We tried the following:

(1) minimax static eval from leaf book positions back up the tree. This failed as there were often much better moves that were overlooked, or a typo was a horrible blunder but was ignored in the analysis.

(2) we went to terminal positions and did searches to make sure that the side that was supposed to be +/- was really winning. This took a lot longer due to the searches and number of terminal positions that needed them. And this still overlooked mid-line blunders and such.

(3) we tried searching at every position from root to tips, so that we could find those bogus things. This provided too much data for us to manage.

(4) we modified that to play a hundred a few very fast games (self-play) at each position, and use that to guide us. That didn't work either as the self-play games were necessarily very fast, which means they overlooked deep tactical issues.

(4) was very similar to what is being described here, and while it offers information, I am not sure how good that information actually is.