A Hessian actor-critic algorithm

We consider Markov Decision Processes (MDPs) following a policy parametrized by a parsimonious set of parameters and seek to optimize the policy over these parameters. In this setting, optimization can be done using a gradient ascent method. If designed well, the parameterized policy can significantly reduce the problem complexity. Existing algorithms usually suffer from slow convergence because they search along the gradient direction in a steepest ascent way. In this paper, we first propose an estimate for the Hessian of the overall reward the decision maker receives. Based on this estimate, we then introduce a new Newton-like method of the actor-critic type. We compare the new algorithm with several existing algorithms in a robotics application and demonstrate that our method exhibits faster convergence.