Dual Representations for Dynamic Programming

• Author(s) / Creator(s)

  • Wang, Tao
  • Bowling, Michael
  • Lizotte, Daniel
  • Schuurmans, Dale

• Technical report TR07-10. We propose to use a new dual approach to dynamic programming. The idea is to maintain an explicit representation of stationary distributions as opposed to value functions. A significant advantage of the dual approach is that it allows one to exploit well developed techniques for representing, approximating, and estimating probability distributions, without running the risks associated with divergent value function estimation. In this paper, we introduce novel dual representations and develop dual algorithms. Moreover, we show that dual algorithms remain stable in situations where standard value function estimation diverges. We also show that the dual view yields a viable alternative to standard value function based techniques and opens new avenues for solving sequential decision making problems. | TRID-ID TR07-10

• Date created

  2007

• Subjects / Keywords

  • Dynamic programming

• Type of Item

  Report

• DOI

  https://doi.org/10.7939/R3959C81D

• License

  Attribution 3.0 International