Reinforcement Learning via AIXI Approximation