Reinforcement learning model: control of nonlinear and unstable processes

Reinforcement learning model: control of nonlinear and unstable processes

Paper in proceedings (conference paper)

Some experiences with using of the reinforcement learning model at control of nonlinear unstable processes are published in this paper. Control process is characterized by extensive depth in such cases, so the learning is computationally very demanding. We propose both using of nonlinear grid of the Q-function approximation table and also using of the learning conception “by expert observation“. Learning off (optimal) control policy is not based on blind searching a state space, but it is in progress with the help of further component, that is able to control the process. Problems are studied on active magnetic bearing one-mass model.

Some experiences with using of the reinforcement learning model at control of nonlinear unstable processes are published in this paper. Control process is characterized by extensive depth in such cases, so the learning is computationally very demanding. We propose both using of nonlinear grid of the Q-function approximation table and also using of the learning conception “by expert observation“. Learning off (optimal) control policy is not based on blind searching a state space, but it is in progress with the help of further component, that is able to control the process. Problems are studied on active magnetic bearing one-mass model.

control algorithm, Q-learning, neural network

control algorithm, Q-learning, neural network

BŘEZINA, T., EHRENBERGER, Z., KRATOCHVÍL, C.

14.06.2001

Institute of Termomechanics Academy of Sience of the Czech Republic, Prague 2001

Svratka

80-85918-64-1

Inženýrská Mechanika 2001

40

2