Minimal Models for Dual Heuristic Programming

S. Vasupongayya, R. Santiago, T. Shannon, and G. Lendaris (USA)


Adaptive Critic, Dual Heuristic Programming, and Jacobian


A differentiable model is required in Dual Heuristic Programming and many other adaptive critic approaches to approximate Dynamic Programming. We use brute force search to find minimal models for two benchmark problems. A minimal model is a minimal set of derivatives that are specified to be non random together with a specification of how non random they need to be. The experimental results show that (1) not all derivatives are required and (2) the required derivatives only need to be qualitatively good with respect to sign most of the time in order for Dual Heuristic Programming to achieve a good controller design. Qualitatively good most of the time means the identified derivatives must be correct in sign mostly while the non-required derivatives can be anything. Results of the type reported here suggest that if an "appropriate" subset of derivatives can be selected for a given dynamical plant, a full system identification process may not be needed for successfully using Dual Heuristic Programming.

Important Links:

Go Back