We consider a problem of learning decision policy from past experience available. Using the Fully Probabilistic Design (FPD) formalism, we propose a new general approach for finding a stochastic policy from the past data. The proposed approach assigns degree of similarity to all of the past closed-loop behaviors. The degree of similarity expresses how close the current decision making task is to a past task. Then it is used by Bayesian estimation to learn an approximate optimal policy, which comprises the best past experience. The approach learns decision policy directly from the data without interacting with any supervisor/expert or using any reinforcement signal. The past experience may consider a decision objective different than the current one. Moreover the past decision policy need not to be optimal with respect to the past objective. We demonstrate our approach on simulated examples and show that the learned policy achieves better performance than optimal FPD policy whenever a mismodeling is present.
