Evolving intelligent embodied agents within a physically accurate environment

This thesis explores the application of evolutionary reinforcement learning techniques for evolving behaviorisms in embodied agents existing within a realistic virtual environment that are subject to the constraints as defined by the Newtonian model of physics. Evolutionary reinforcement learning uses evolutionary computation techniques, which are based to some degree, on the evolution of biological life in the natural world. These techniques are generally stochastic in nature and involve random decisions that guide the optimization process via processes of selection, mutation and reproduction. A common problem of using evolutionary computation techniques to evolve intelligent behaviors in embodied agents is the simplicity of the environment and overall system often precludes any life-like behaviors from emerging. Furthermore, the commonly used supervised learning techniques are extremely difficult to apply to embodied agents that employ a complex control system. This thesis proposes a methodology, based on neuroevolution, that effectively addresses this issue of environmental complexity and learning; thus, allowing for the emergence of life-like and efficient behaviors.