No abstract

In this paper, a novel method for personal intelligent behavior avatar (IBA) is proposed to acquire autonomous behavior based on the interactions between user and smart objects in the virtual environment. In this method, the behavior decision model and the self-learning model are integrated by Bayesian Networks and reinforcement learning. The Bayesian Networks can treat interaction experiences using statistical processes, and the sureness of decision making is represented by certainty factors using stochastic reasoning. The reinforcement learning is implemented by learning experimentation or trial and error mechanisms to improve the performance of IBA through feedback. Therefore, the IBA makes a strategic decision that is approximated and appropriate to the user through the self-learning process by reinforcement learning. Finally, the feasibility of this method is investigated by imitating user's behavior and the results of self-learning process. The results of simulation show that the method is successful in imitating user's behavior and improving the performance of IBA..

The research topic of this paper is that how an avatar explores the environment, and find the way out to reach the pre-assigned goal. This proposed system is applied to a dynamicallly changable, continuous and largescaled environment. To deal with the learning in such environment, the proposed system partitions the statespace into regions of states, called cells. By using cells, the states of the system can be reduced in which the number of states grows dramatically increasing in proportion to the number and quality of inputs. Through a series of maniputations of cells, the avatar can adapt itself to the changable environment. By adjusting the peferences of the proposed global-cell, the tendancy of exploration behavior of the avatar can be controlled to explore the potential path caused by the changing environment. KEYWORDSvirtual world, dynamic environment, reinforcement learning, machine learning, curse of dimentionality.