Evolving Fault Tolerant Robotic Controllers

Fault tolerant control and evolutionary algorithms are two different research areas. However with the development of artificial intelligence, evolutionary algorithms have demonstrated competitive performance compared to traditional approaches for the optimisation task. For this reason, the combination of fault tolerant control and evolutionary algorithms has become a new research topic with the evolving of controllers so as to achieve different fault tolerant control schemes.
However most of the controller evolution tasks are based on the optimisation of controller parameters so as to achieve the fault tolerant control, so structure optimisation based evolutionary algorithm approaches have not been investigated as the same level as parameter optimisation approaches. For this reason, this thesis investigates whether structure optimisation based evolutionary algorithm approaches could be implemented into a robot sensor fault tolerant control scheme based on the phototaxis task in addition to just parameter optimisation, and explores whether controller structure optimisation could demonstrate potential benefit in a greater degree than just controller parameter optimisation.
This thesis presents a new multi-objective optimisation algorithm in the structure optimisation level called Multi-objective Cartesian Genetic Programming, which is created based on Cartesian Genetic Programming and Non-dominated Sorting Genetic Algorithm 2, in terms of NeuroEvolution based robotic controller optimisation. In order to solve two main problems during the algorithm development, this thesis investigates the benefit of genetic redundancy as well as preserving neutral genetic drift in order to solve the random neighbour pick problem during crowding fill for survival selection and investigates how hyper-volume indicator is employed to measure the multi-objective optimisation algorithm performance in order to assess the convergence for Multi-objective Cartesian Genetic Programming.
Furthermore, this thesis compares Multi-objective Cartesian Genetic Programming with Non-dominated Sorting Genetic Algorithm 2 for their evolution performance and investigates how Multi-objective Cartesian Genetic Programming could be performing for a more difficult fault tolerant control scenario besides the basic one, which further demonstrates the benefit of utilising structure optimisation based evolutionary algorithm approach for robotic fault tolerant control.