Output details
11 - Computer Science and Informatics
University of Edinburgh
Expectation-Maximization Methods for Solving (PO)MDPs and Optimal Control Problems.
<24> Originality: This was the first work to establish a general framework for using inference technology for solving general partially observable Markov Decision Processes with general horizons.
Significance: The paper establishes direct equivalence between control optimization and expectation maximization (EM) for inference, and thereby enabling the use of EM approaches for solving POMDPs. EM is now a popular approach for such problems as it is efficient due to the decoupling of forward and backward passes.
Rigour: The paper establishes direct mathematical equivalence between control optimization and expectation maximization for inference.