Model-Free Reinforcement Learning as Mixture Learning

We cast model-free reinforcement learning as the problem of maximizing the likelihood of a probabilistic mixture model via sampling, addressing both the innite and nite horizon cases. We describe a