Information-Theoretic Methods for Planning and Learning in Partially Observable Markov Decision Processes

@inproceedings{Tishby2016InformationTheoreticMF,
  title={Information-Theoretic Methods for Planning and Learning in Partially Observable Markov Decision Processes},
  author={Naftali Tishby},
  year={2016}
}
We model the interaction of an intelligent agent with its environment as a Partially Observable Markov Decision Process (POMDP), where the joint dynamics of the internal state of the agent and the external state of the world are subject to extrinsic and intrinsic constraints. Extrinsic constraints of partial observability and partial controllability specify how the agent’s input observation depends on the world state and how the latter depends on the agent’s output action. The agent also incurs… CONTINUE READING