Bio-inspired Dynamic Control Systems with Time Delays

Abstract

The world around us exhibits a rich and ever changing environment of startling, bewildering and fascinating complexity. Almost everything is never as simple as it seems, but through the chaos we may catch fleeting glimpses of the mechanisms within. Throughout the history of human endeavour we have mimicked nature to harness it for our own ends. Our attempts to develop truly autonomous and intelligent machines have however struggled with the limitations of our human ability. This has encouraged some to shirk this responsibility and instead model biological processes and systems to do it for us.

This Thesis explores the introduction of continuous time delays into biologically inspired dynamic control systems. We seek to exploit rich temporal dynamics found in physical and biological systems for modelling complex or adaptive behaviour through the artificial evolution of networks to control robots. Throughout, arguments have been presented for the modelling of delays not only to better represent key facets of physical and biological systems, but to increase the computational potential of such systems for the synthesis of control.

The thorough investigation of the dynamics of small delayed networks with a wide range of time delays has been undertaken, with a detailed mathematical description of the fixed points of the system and possible oscillatory modes developed to fully describe the behaviour of a single node. Exploration of the behaviour for even small delayed networks illustrates the range of complex behaviour possible and guides the development of interesting solutions.

To further exploit the potential of the rich dynamics in such systems, a novel approach to the 3D simulation of locomotory robots has been developed focussing on minimising the computational cost. To verify this simulation tool a simple quadruped robot was developed and the motion of the robot when undergoing a manually designed gait evaluated. The results displayed a high degree of agreement between the simulation and laser tracker data, verifying the accuracy of the model developed.

A new model of a dynamic system which includes continuous time delays has been introduced, and its utility demonstrated in the evolution of networks for the solution of simple learning behaviours. A range of methods has been developed for determining the time delays, including the novel concept of representing the time delays as related to the distance between nodes in a spatial representation of the network. The application of these tools to a range of examples has been explored, from Gene Regulatory Networks (GRNs) to robot control and neural networks. The performance of these systems has been compared and contrasted with the efficacy of evolutionary runs for the same task over the whole range of network and delay types.

It has been shown that delayed dynamic neural systems are at least as capable as traditional Continuous Time Recurrent Neural Networks (CTRNNs) and show significant performance improvements in the control of robot gaits. Experiments in adaptive behaviour, where there is not such a direct link between the enhanced system dynamics and performance, showed no such discernible improvement. Whilst we hypothesise that the ability of such delayed networks to generate switched pattern generating nodes may be useful in Evolutionary Robotics (ER) this was not borne out here.

The spatial representation of delays was shown to be more efficient for larger networks, however these techniques restricted the search to lower complexity solutions or led to a significant falloff as the network structure becomes more complex. This would suggest that for anything other than a simple genotype, the direct method for encoding delays is likely most appropriate. With proven benefits for robot locomotion and the open potential for adaptive behaviour delayed dynamic systems for evolved control remain an interesting and promising field in complex systems research.