Monte-Carlo Sampling in Games with Imperfect Information: Empirical Investigation and Analysis

We investigate Monte-carlo sampling in games with imperfect information. We show that for very simple game trees the chance of nding the optimal strategy with Monte-carlo sampling rapidly approaches zero as the number of moves in the game increases. We explain this sub-optimality by identifying the diierent kinds of errors that can arise, and by analysing their interplay. We also relate our test results to real games, suggesting why the error rates observed in practice may not be so high.