Binary action search for learning continuous-action control policies