1994 Volume 7 Issue 12 Pages 487-497
The control of a group of elevators is a difficult stochastic control problem, because of the random and unpredictable passenger arrivals. Here we propose a new method for constructing an adaptive controller for stochastic systems, by using a combination of neural networks and an on-line reinforcement learning based on stochastic approximation. This method combines an efficient supervised learning with a general reinforcement adaptation. The supervised learning is used to prepare the controller with domain-specific knowledge, and for initializing it to emulate an existing controller. The reinforcement learning is used for adaptation, with a special attention paid to allow online operation. The new method is used to develop an adaptive, optimal elevator group controller. Results of simulation tests indicate the feasibility of using the proposed method for industrial applications.