A menu of designs for reinforcement learning over time
Neural networks for control
Neuro-Dynamic Programming
| Hi-index | 0.00 |
This paper presents the results of simulation and control experiments using a recently proposed method for real-time switching among a pool of controllers. The switching strategy selects the current controller based on neural network estimates of the future system performance for each controller. This neural-network-based switching controller has been implemented for a simulated inverted pendulum and a level control system for an underwater vehicle in our laboratory. The objectives of the experiments presented here are to demonstrate the feasibility of this approach to switching control for real systems and to identify techniques to deal with practical issues that arise in the training of the neural networks and the real-time switching behavior of the system. This experimental work complements on-going theoretical investigations of the method which will be reported elsewhere.