An Effective Rule Based Policy Representation and its Optimization using Inter Normal Distribution Crossover

GA has an advantage that it can treat target functions without depending on their forms. Recently, many studies have been conducted on applying GA into the policy search. Especially, approaches using a rule based policy with Gaussian Mixture are promising. There is not, however, any genetic operator to create a new normal distribution from plural ones. We propose an effective policy representation and a new genetic operator INDX for it. The performance of the proposed method is shown, applying it to two benchmark problems, the Mountain Car and the Cart-Pole Swing up task.