特邀嘉宾---Assoc. Prof. Daisuke Kitakoshi


Division of Computer Science, National Institute of Technology, Tokyo College, Japan


Biography: Daisuke Kitakoshi received the BEng in Computer Science, the MSc (IT), and the Ph. D. degrees in Systems and Information Engineering from Hokkaido University, Japan, in 1998, 2000, and 2003. He is an associate professor at Division of Computer Science, National Institute of Technology, Tokyo College, Japan since 2011, after having contributed as a lecturer at National Institute of Technology, Tokyo College since Jun. 2008, and having contributed as an assistant professor to the Department of Scientific and Engineering Simulation (since Apr. 2008) and the Department of Computer Science and Engineering (since 2004), Graduate School of Engineering at Nagoya Institute of Technology, Japan.
His current research interests include Artificial Intelligence, Machine Learning, Data Mining, and Human-Agent (Robot) Interaction. He is also engaging in research on hybrid systems that contribute to adaptive behavior of agents in dynamic environments.

Speech Title: Hybrid Adaptive Learning System based on Reinforcement Learning and Mixture Probability
Abstract: In a variety of fields such as disaster relief, automated cruise, and RoboCup, there has been increasing attention to learning objects such as autonomous robots and software agents which can adaptively behave to attain their goals in dynamic environments. Although many research efforts have been devoted to "flexible adaptation" of the learning objects to changes in complicated environments, there still remained problems in terms of their adaptive ability and/or computational complexity. This talk will present an on-line system, inspired by the human ability to act based on empirical knowledge, that allows reinforcement learning agents to adapt to environmental changes by using a mixture of stochastic models, or probabilistic distributions. Distinct mixture formulation called exponential mixture is incorporated into the system to improve its adaptive performance. Clustering component of the mixture also contributes to reduction of computational complexity while maintaining adaptability. We present several results that mobile robots and agents with the proposed adaptive learning system behaved in a flexible manner in both complicated virtual environments and real-world environments, and show that the concept of exploiting knowledge expressed by mixture probability is promising for applying to various practical fields.