A performance gradient perspective on approximate dynamic programming and its application to partially observable Markov decision processes

This paper shows an approach to integrating common approximate dynamic programming (ADP) algorithms into a theoretical framework to address both analytical characteristics and algorithmic features. Several important insights are gained from this analysis, including new approaches to the creation of algorithms. Built on this paradigm, ADP learning algorithms are further developed to address a broader class of problems: optimization with partial observability. This framework is based on an average cost formulation which makes use of the concepts of differential costs and performance gradients to describe learning and optimization algorithms. Numerical simulations are conducted including a queueing problem and a maze problem to illustrate and verify features of the proposed algorithms. Pathways for applying this analysis to adaptive critics are also shown.