Exploration of Neural Structures for Dynamic System Control

Biological neural systems are powerful mechanisms for controlling biological sys- tems. While the complexity of biological neural networks makes exact simulation intractable, several key aspects lend themselves to implementation on computational systems.

This thesis constructs a discrete event neural network simulation that implements aspects of biological neural networks. A combined genetic programming/simulated annealing approach is utilized to design network structures that function as regulators for continuous time dynamic systems in the presence of process noise when simulated using a discrete event neural simulation.

Methods of constructing such networks are analyzed including examination of the final network structure and the algorithm used to construct the networks. The parameters of the network simulation are also analyzed, as well as the interface between the network and the dynamic system. This analysis provides insight to the construction of networks for more complicated control applications.