In Reinforcement Learning, epsilon-greedy policies are the most used exploration policies, but in case there is a big state space with impossible actions, wouldn't it be better to use soft-max policies instead?