Recurrent neural networks and adaptive motor control

Abstract

This thesis is concerned with the use of neural networks for motor control tasks. The main goal of the thesis is to investigate ways in which the biological notions of motor programs and Central Pattern Generators (CPGs) may be implemented in a neural network framework. Biological CPGs can be seen as components within a larger control scheme, which is basically modular in design. In this thesis, these ideas are investigated through the use of modular recurrent networks, which are used in a variety of control tasks. The first experimental chapter deals with learning in recurrent networks, and it is shown that CPGs may be easily implemented using the machinery of backpropagation. The use of these CPGs can aid the learning of pattern generation tasks; they can also mean that the other components in the system can be reduced in complexity, say, to a purely feedforward network. It is also shown that incremental learning, or 'shaping' is an effective method for building CPGs. Genetic algorithms are also used to build CPGs; although computational effort prevents this from being a practical method, it does show that GAs are capable of optimising systems that operate in the context of a larger scheme. One interesting result from the GA is that optimal CPGs tend to have unstable dynamics, which may have implications for building modular neural controllers. The next chapter applies these ideas to some simple control tasks involving a highly redundant simulated robot arm. It was shown that it is relatively straightforward to build CPGs that represent elements of pattern generation, constraint satisfaction. and local feedback. This is indirect control, in which errors are backpropagated through a plant model, as well as the ePG itself, to give errors for the controller. Finally, the third experimental chapter takes an alternative approach, and uses direct control methods, such as reinforcement learning. In reinforcement learning, controller outputs have unmodelled effects; this allows us to build complex control systems, where outputs modulate the couplings between sets of dynamic systems. This was shown for a simple case, involving a system of coupled oscillators. A second set of experiments investigates the use of simplified models of behaviour; this is a reduced form of supervised learning, and the use of such models in control is discussed.