Reinforcement learning with knowledge by using a stochastic gradient method on a Bayesian network

For real applications of reinforcement learning, it is necessary to reduce the number of trial-and-errors. The paper proposes a method to use knowledge in reinforcement learning. We have regarded a Bayesian network as a stochastic policy, and adapted a rigid propagation procedure for a stochastic gradient method. We made preliminary experiments to demonstrate our method in a robot navigation task.