## Future Information Minimization as PAC Bayes regularization in Reinforcement Learning

author: Naftali Tishby, School of Computer Science and Engineering, The Hebrew University of Jerusalem
published: Jan. 25, 2012,   recorded: December 2011,   views: 156
Categories
You might be experiencing some problems with Your Video player.

# Slides

0:00 Slides Future Information Minimization as PAC Bayes regularization in Reinforcement Learning [Partially Observed] Markov Decision Processes Reinforcement Learning revisited The Agent Learns a Policy Graphical model for the perception-action-cycle Information pickup: I-gains and Bellman optimality Bellman meets Shannon Decision-sequences and information Decision/action sequences and information Proof idea: Recursive Information-chain rules ... Application: Uncertainty reduction and source coding Application: Huffman coding and Bellman optimality Application: Sequential hypothesis testing How much information is needed for valuable behavior? Value (extrinsic) and Information (intrinsic) ... Value (extrinsic) and Information (intrinsic) ... Combining (future) Value and Information Trading Value and (future) Information Information bounded RL Maze More complex maze ... - 1 Animation - 1 Animation - 2 Animation - 3 More complex maze ... - 1 More complex maze ... - 2 Global convergence theorem PAC-Bayes Generalization Theorem (McAllester 2001) PAC-Bayes Robustness Theorem for I-RL The optimal tradeoff between Value and Future Information Beyond MDP: Extracting only valuable past information Sequential Information gathering Conclusions

# Report a problem or upload files

If you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.
Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.

# Description

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.