| | |
Summary: Reinforcement Learning with Modular Neural Networks for Control
Charles W. Anderson Zhaohui Hong
Department of Computer Science
Colorado State University
Fort Collins, CO 80523
anderson@cs.colostate.edu
Abstract
Reinforcement learning methods can be applied to
control problems with the objective of optimizing the
value of a function over time. They have been used
to train single neural networks that learn solutions to
whole tasks. Jacobs and Jordan 5 have shown that a
set of expert networks combined via a gating network
can more quickly learn tasks that can be decomposed.
Even the decomposition can be learned. Inspired by
Boyan's work of modular neural networks for learning
with temporal-di erence methods 4 , we modify the
reinforcement learning algorithm called Q-Learning to
train a modular neural network to solve a control prob-
lem. The resulting algorithm is demonstrated on the
|
