IGW

An exploration strategy that allocates selection probabilities according to the currently estimated action gaps, exploring more when gaps are smaller. It avoids explicit optimization of information gain, was originally used in bandit settings, and was later generalized to partial monitoring.