An Alternative to Backpropagation in Deep Reinforcement Learning