b) When training an agent to play a two-player game against other agents trained by other people (e.g., chess, rock-paper-scissors), it might be useful to use a stochastic policy both at train and test time