Hi Dave,

> Developing a UCT robot for a new game, I have encountered a
> surprising and alarming behavior:  the longer think time the
> robot is given, the worse the results.  

Can you tell us the rules of the game?
Maybe they help to explain the phenomenon.
(Once, Cameron Prowne had strange MC behavior in one of his newly
invented games.)

A few years ago, Wesley M. Turner and I artificially designed (simple)
games with the intention to achieve strange MC-bot behavior.
We were very successful. Our findings are in a paper
"Anomalies of pure monte-carlo search in monte-carlo perfect games"

"MC-perfect game" means that the performance of pure MC (without tree
part) converges to optimal play when playout number goes to infinity.
We were able to construct such games where for instance MC(1000)
achieves less than 1 percent of wins in play against MC(1).

Here, MC(k) is the agent that plays k random games for each
succesor of the current position.

Our paper is in the proceedings of
Computers and Games: 8th International Conference, CG 2013, Yokohama, Japan.
Editors H.J. van den Herik, H. Iida, A. Plaat.
Springer Lectures Notes in Computer Science 8427, p.84-99.

Computer-go mailing list

Reply via email to