The Dangers of Random Playouts

With recent advances in Monte Carlo approaches to move planning in computer games – in particular Monte Carlo Tree Search (MCTS) proposed by Coulomb (2006) – more attention is now being paid to simulationbased methods such as the use of random playouts to estimate the utility of given board positions. The general assumption is that the accuracy of these estimates will increase with the number of simulations; we present a simple counterexample that demonstrates that this is not always the case.