Simultaneous incremental neuroevolution of motor control, navigation and object manipulation in 3D virtual creatures

Abstract

There have been numerous attempts to develop 3D virtual agents by applying evolutionary processes to populations that exist in a realistic physical simulation. Whilst often contributing useful knowledge, no previous work has demonstrated the capacity to evolve a sequence of increasingly complex behaviours in a single, unified system. This thesis has this demonstration as its primary aim. A rigorous exploration of one aspect of incremental artificial evolution was carried out to understand how subtask presentations affect the whole-task generalisation performance of evolved, fixed-morphology 3D agents. Results from this work led to the design of an environment–body–control architecture that can be used
as a base for evolving multiple behaviours incrementally. A simulation based on this architecture with a more complex environment was then developed and explored. This system was then adapted to include elements of physical
manipulation as a first step toward a fully physical virtual creature environment demonstrating advanced evolved behaviours.

The thesis demonstrates that incremental evolutionary systems can be subject to problems of forgetting and loss of gradient, and that different complexification strategies have a strong bearing on the management of these issues. Presenting successive generations of the population to a full range of objective functions (covering and revisiting the range of complexity) outperforms straightforward linear or direct presentations, establishing a more robust approach to the evolution of naturalistic embodied agents. When combining this approach with a bespoke control architecture in a problem requiring reactive and deliberative behaviours, we see results that not only demonstrate success at the tasks, but also show a variety of intricate behaviours being used. This is the first ever example of the simultaneous incremental evolution in 3D of composite behaviours more complex than simple locomotion. Finally, the architecture demonstrably supports extension to manipulation in a feedback control task. Given the problem-agnostic controller architecture, these results indicate a system with potential for discovering yet more advanced behaviours in yet more complex environments.