High fidelity virtual environments can be inhabited by virtual creatures. A virtual creature should be able to learn itself how to improve its old behaviors and produce new related behaviors so as to be more adaptive and autonomous and hence reduce human design work. This thesis presents a study of an adaptive and multifunctional Computational Behavior Generation (CBG) model for virtual creatures with the ultimate goal of enhancing a creature's adaptation and multifunctionality in behavior control by learning. Specifically, we require that the CBG model can learn to perform variable behavior tasks in various environments and situations. The design of the CBG model is inspired by the natural behavior control system in the brain. It can perform the whole procedure of decision, programming and execution of motor actions, and its hierarchical architecture provides the material basis for its adaptive and multifunctional learning implementation. The concrete achievement of adaptation and multifunctionality by learning is obtained with the help of a Multiagent based Evolutionary Artificial Neural Network with Lifetime Learning (MENL), which can learn to make correct action decisions for varied behavior tasks in varied situations. MENL takes advantage of the whole population information of evolution by maintaining a batch of multiagents in every evolutionary generation. These agents co-decide the decisions to be executed, and they are subject to evolutionary learning through all of their lifetime. The fitness function of MENL is designed without many specific constraints, and can be easily extended for a variety of behaviors. As a consequence, the CBG with MENL can obtain high adaptation and generalization in behavior. The CBG model combined with the MENL learning algorithm enables a virtual creature to learn several general navigation functions independently and jointly in unknown environments. These functions include exploration, goal reaching, and wandering. The virtual creature is first asked to learn general exploration only in a series of increasing complex environments. This creature has adapted to various environments and navigated in them successfully. The whole successful exploration experiment is achieved due to the competition and emergent cooperation among multiagents and their continuous lifetime learning.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:663446 |
| Date | January 2002 |
| Creators | Wang, Fang |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/13189 |
Page generated in 0.0023 seconds