Evolving virtual agents using genetic programming

Abstract

Virtual environments are used in a diverse number of applications, ranging from medical applications, military simulations, modeling and engineering, to entertainment such as games or virtual communities. In these applications, virtual agents can be used to make the environment more realistic, perform tasks that are tedious and time consuming for humans, or even simulate the presence of other users in the environment. When constructing an agent for a virtual environment, several issues are encountered that must be resolved. First, a virtual agent must be able to explore and navigate in the virtual environment in a realistic way while avoiding collisions with obstacles. If the virtual agent does not have access to the internal representation of the environment, it will have to use its virtual sensors to observe the environment. In this thesis, an algorithm is presented to perform obstacle avoidance and map construction in a virtual environment using a synthetic vision sensor. The constructed map can then also be used to navigate in the environment. A second issue is communication between agents and users in the environment. Agents and users must be able to Locate agents that can perform certain tasks, and agents may offer their services to users or other agents. These issues are discussed briefly in this thesis, and a prototype of a multi-agent virtual environment is presented. The most difficult issue of virtual agents is learning to solve problems in an environment, without knowing the constraints and rules of the environment in advance. This thesis will examine the use of genetic programming to train virtual agents. Two important problems are encountered when using genetic programming in this domain. First, programs constructed using genetic programming tend to grow rapidly before an acceptable solution is found. Several techniques will be presented to reduce the size of the evolved genetic programs, and a comparison will be made between these techniques. Secondly, evaluation of candidate solutions is usually very time consuming, making it impractical to maintain a large population of candidate solution. A large population is usually a requirement to evolve good solutions. Therefore, an algorithm to reduce the size of the population while maintaining the diversity of a larger population is presented. These optimizations will also be applied to the virtual multi-agent system of robotic soccer to examine the effects of these optimizations in a complex environment.