Monte Carlo go 囲碁 programs currently play "slack" moves if they are sure they are going to win, or lose, which is aesthetically not pleasing. This could easily be fixed by adjusting komi when it starts doing that (which can be detected when the evaluation is near 0 or 1). Evaluate the position for a range of komi. This of course is computationally more expensive, maybe doable for analysis.
Just how strongly can Monte Carlo go 囲碁 programs play endgames? As well as a human pro? Better? Can they handle combinatorial game theory as described in Mathematical Go?
No comments :
Post a Comment