Evolution of Multiobjective Neuromodulated Neurocontrollers for Multi-Robot Systems

Abstract

This thesis focuses on the evolution of multiobjective neurocontrollers with the ability to perform unsupervised learning while operating. We begin with some biologically-inspired modifications to a standard neuroevolution algorithm. Synaptic plasticity has been shown to facilitate unsupervised learning by adapting neural network weights. Neuromodulation is a technique that can adapt the per-connection learning rates of synaptic plasticity. Multiobjective evolution of neural network topology and weights has been used to design autonomous neurocontrollers. Multiobjective evolution of network weights and topologies (NEAT-MODS) is augmented with neuromodulated learning. It is also shown that speciation is unnecessary when neuromodulation is used. The design is demonstrated using a series of experiments with simulated robots navigating a maze containing goals. It is shown that when neuromodulated learning is combined with evolution, neurocontrollers are synthesized in fewer generations than by evolution alone. Secondly, Lamarckian inheritance (passing learned behaviour from parent to offspring generations) is used with the neuromodulated non-speciated NEAT-MODS method. The Lamarckian-inherited neuromodulated approach is found to be statistically superior to neuromodulated NEAT-MODS, and NEAT-MODS alone when applied to solve a multiobjective navigation problem. Six new contributions are presented. Firstly, the two previous investigations are demonstrated using pursuit-evasion games which are more complex than the previously demonstrated foraging maze task. Secondly, effectiveness of the neurocontrollers when applied to robots in the pursuit-evasion game that are modeled as inertial elements that can undergo limited acceleration is demonstrated. Thirdly, an investigation into the performance of composite and atomic objectives on neuromodulated NEAT-MODS with Lamarckian inheritance controlling the behaviour of the robots in the pursuit-evasion game is performed. Fourthly, the evolutionary algorithm is augmented to include all system parameters in order to reduce reliance on operator tuning methods that are commonly used to optimize these values for a given problem. The fifth contribution involves examining the effectiveness of implementing evolution with objective hierarchy using the novel Objective Weighted Ranking (OWR) technique. OWR prioritizes the evolution of fitness in a fundamental objective reducing the time required to evolve suitable candidates. The sixth contribution investigates the relationship between neurocontroller topology and function, to explain the relationship between neural-network structure and behaviour.