Statistics of Policy search in a space of simple closed-form formulas: towards interpretability of reinforcement learning