W. Tang, C. Gatzoulis, and T.R. Wan (UK)
virtual reality, intelligent virtual agent, reinforcement learning, and artificial intelligence
This paper describes our investigation in ethology approaches to build proactive anticipatory virtual characters. The control architecture of the characters is based on two Reinforcement Learning Algorithms: the zeroth-level Classifier system (ZCS) and the Q-Learning Algorithm. The ZCS consists of inter-connected functional components, each of which is a container of classifiers that evolve through interactions with the environment, and the multiplicity and recursion of the interconnected sub-units of the control system. In the Q-Learning control system, the Q algorithm is combined with eligibility traces, a temporary reward assignment mechanism. Both systems are tested and the results are compared in this framework to investigate the complexity of Reinforcement Learning algorithms and their application in the creation of proactive virtual characters that have internal needs and can discover optimal ways of achieving a predefined goal through the evolutionary learning process.
Important Links:
Go Back
