Predictive Hebbian learning of representation in a fast reinforcement controller

Author

Schultz, Simon R. ; Jabri, Marwan A.

Author_Institution

Dept. of Electr. Eng., Sydney Univ., NSW, Australia

fYear

1994

fDate

29 Nov-2 Dec 1994

Firstpage

56

Lastpage

60

Abstract

A particular version of the cart-pole problem has recently been solved trivially by Moody and Tresp (1994). We present a reinforcement learning pole balancer which learns a solution to the problem nearly as quickly. Our controller, however, does not make use of a predefined representation of the input space, but instead learns an appropriate representation using a multilayer evaluation network. The hidden (representation) layer of the evaluation network is adapted using a predictive Hebbian learning algorithm

Keywords

Hebbian learning; adaptive control; feedforward neural nets; multilayer perceptrons; multivariable control systems; nonlinear control systems; predictive control; cart-pole problem; fast reinforcement controller; hidden layer adaptation; multilayer evaluation network; pole balancer; predictive Hebbian learning algorithm; reinforcement learning; representation learning; Automatic control; Control systems; Design automation; Design engineering; Hebbian theory; Laboratories; Learning; Neurons; Prediction algorithms; Systems engineering and theory;

fLanguage

English

Publisher

ieee

Conference_Titel

Intelligent Information Systems,1994. Proceedings of the 1994 Second Australian and New Zealand Conference on

Conference_Location

Brisbane, Qld.

Print_ISBN

0-7803-2404-8

Type

conf

DOI

10.1109/ANZIIS.1994.396950

Filename

396950