REINFORCEMENT LEARNING CONTROL SCHEMES

Vilma Alves de Oliveira, Eduardo Fontoura Costa, Aluízio Fausto Ribeiro Araújo e Renato Tinós

Departamento de Engenharia Elétrica
Universidade de São Paulo
Caixa Postal 359, 13560-970, São Carlos, SP, Brazil
e-mail: vilmao@sel.eesc.sc.usp.br

Resumo:

Abstract:

In this paper, the use of Artificial Neural Network for control of non-linear plants is explored. As the plant parameters or model is considered unknown, it is necessary to use plant input/output to train the controller and therefore reinforcement learning control schemes are devised to achieve desired results. The main features of the developed schemes are that few trials are required to train the controllers and a variety of control actions is taken rather than only two actions as in the standard reinforcement schemes. In addition, a supervised neural network controller, which is trained using the reinforcement control schemes, is proposed. An example of a magnetic suspension system is presented to illustrate the effectiveness of the control algorithms given. For comparison purposes, results of a linear optimal controller are included.

Key Words: Intelligent controllers, Adaptive critics, Reinforcement learning, Neural networks

Cópia em PDF (136 kbytes)

Volta ao Índice do Volume 9 .