Consider non-stationary Bandit algorithm for adaptive mutations

[gkirgizov]

Distribution of which mutation is best is in general non-stationary during optimization process. So we need considering appropriate algorithms, because default bandit algorithms assume stationary distribution p(reward|action).