SJE's posts got me thinking.
How strong a correlation there would be between a monte carlo playout of a position vs. a traditional search? It seems that monte carlo would perform poorly when there is a narrow forced win. But perhaps the reverse is true, and random playout could identify positions where an engine's evaluation has a blind spot.
Has anyone done anything like that?
Random playout vs evaluation
Moderators: hgm, Rebel, chrisw
-
- Posts: 12550
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Re: Random playout vs evaluation
Monte Carlo search is used all the time in go.
I don't know if anyone has tried it in chess.
https://chessprogramming.wikispaces.com ... ree+Search
Here is a github monte carlo tree search:
https://github.com/lteacy/mcts-cpp
I don't know if anyone has tried it in chess.
https://chessprogramming.wikispaces.com ... ree+Search
Here is a github monte carlo tree search:
https://github.com/lteacy/mcts-cpp
-
- Posts: 931
- Joined: Tue Mar 09, 2010 3:46 pm
- Location: New York
- Full name: Álvaro Begué (RuyDos)
Re: Random playout vs evaluation
I don't expect Monte-Carlo simulations would work very well in chess. Certainly not when giving every move the same probability ("light playouts", in MCTS lingo).
If you write something a bit smarter, so you don't drop pieces for no reason, perhaps you get some reasonable results. I would be curious to see a MCTS chess engine that uses depth-1 search during the playouts.
If you write something a bit smarter, so you don't drop pieces for no reason, perhaps you get some reasonable results. I would be curious to see a MCTS chess engine that uses depth-1 search during the playouts.
-
- Posts: 4185
- Joined: Tue Mar 14, 2006 11:34 am
- Location: Ethiopia
Re: Random playout vs evaluation
I tried this a while ago for chess and it sucked. MCTS works awesomely for checkers though because there captures are forced unlike in chess. And that seems to make a big difference more than anything else to the accuracy of the random playout. Nebiyu has MCTS mode for playing many types of games if anyone is interested.AlvaroBegue wrote:I don't expect Monte-Carlo simulations would work very well in chess. Certainly not when giving every move the same probability ("light playouts", in MCTS lingo).
If you write something a bit smarter, so you don't drop pieces for no reason, perhaps you get some reasonable results. I would be curious to see a MCTS chess engine that uses depth-1 search during the playouts.
Re: Random playout vs evaluation
I tried it once, in a distributed version (+/- 120 computers).
Failed
Failed