Sequential Monte Carlo samplers for model-based reinforcement learning

Reinforcement learning problems are generally solved by using fixed-point iterations that converge to the suboptimal solutions of Bellman equations. However, it is also possible to formalize this problem as an equivalent likelihood maximization problem and employ probabilistic inference methods. We proposed an expectation-maximization algorithm that utilizes sequential Monte Carlo samplers with Metropolis-Hastings kernels in its expectation step to solve the model-based version. Then, we evaluate our algorithm on mountain-car problem which is a benchmark reinforcement learning problem.