Interaction-limited policy learning