Future Information Minimization as PAC Bayes regularization in Reinforcement Learning

Interactions between an organism and its environment are commonly treated in the framework of Markov Decision Processes (MDP). While standard MDP is aimed at maximizing expected future rewards (value), the circular flow of information between the agent and its environment is generally ignored. In particular, the information gained from the environment by means of perception and the information involved in the process of action selection are not treated in the standard MDP setting. In this talk, we focus on the control information and show how it can be combined with the reward measure in a unified way. Both of these measures satisfy the familiar Bellman recursive equations, and their linear combination (the free-energy) provides an interesting new optimization criterion. The tradeoff between value and information, explored using our INFO-RL algorithm, provides a principled justification for stochastic (soft) policies. These optimal policies are also shown to be robust to uncertainties in the reward values by applying the PAC-Bayes generalization bound. The same PAC-Bayesian bounding term thus plays the dual roles of information-gain in the Information-RL formalism and as a model-order regularization term in the learning of the process.