97-13   Berichtsreihe des Mathematischen Seminars der Universität Kiel

Stephan Pareigis, Martin Riedmiller:

A hybrid grid refinement scheme for reinforcement learning based on local defect correcting methods

We present an adaptive grid scheme for reinforcement learning. Initially, the value function is approximated on a coarse fixed cell-centered grid. A residual based indicator function for the approximation error is derived. It determines critical regions, in which the grid needs to be refined. A neural net is then trained to approximate the value function of the local problem in the critical region. In order to guarantee a consistent interaction between the two levels of approximation, a method from numerical analysis, the local defect-correction (LDC) method is adapted. A simple numerical experiment shows the feasibility and the working principles of the proposed adaptive grid scheme method.

Keywords: reinforcement learning, dynamic programming, adaptive grid refinement, local defect-correction methods, neural networks.

Mail an Jens Burmeister
[Thu Feb 19 18:56:33 2009]
Impressum