Intelligent Systems in Nanjing University

Intelligent systems is a major research theme in Nanjing University, with the support from the State Key Laboratory for Novel Software Technology of China, one of the top laboratories in the information technology field in the whole country. Currently, the research carried out by the intelligent systems group at Nanjing University mainly fo-cuses on the following topics: • Fundamental methods of intelligent computing, particularly reinforcement learning, incremental learning , granular computing and rough sets. • Intelligent agents and multi-agent systems. • Content-based multimedia (images and 3D models) retrieval. The intelligent systems group aims to design intelligent algorithms and systems to deal with real problems efficiently. The group has developed software tools and applications covering wide areas, including medical treatment, public security, and virtual games. • Reinforcement learning Reinforcement learning is an on-line, incremental learning technology, by which intelligent agents interact with the surrounding world by trial-and-error, and learn the optimal policy of decision sequences according to reinforcement signals. Our group has studied various algorithms for reinforcement learning problems, including average reward reinforcement learning, multi-agent reinforcement learning, relational reinforcement learning, function approximation in reinforcement learning and option discovery , etc. For example, our G-learning algorithm addresses the average reward domain, and is more stable than the classical R-learning and Q-learning. The following figure shows four curves of the average rewards computed in every 10,000 steps in the access-control queuing task. As shown in this figure, G-learning outperforms R-learning and its variation in the learning speed. Besides the theoretical algorithm research , the group also studies the application of reinforcement learning. The proposed algorithms have been applied to video game tasks such as Tetris. The group attended the RL 2008 competition and won the 6th place in the game of Tetris. Research in learning classi-fier systems is related to reinforcement learning as well. Our group has successfully applied the learning classifier systems technology to different problem domains including data mining, medical data analysis and image steganography detection. • Rough set Rough set theory is a sound mathematical tool to deal with imprecise, uncertain, and vague information. Most of the group's work in this area can be characterized by rough-set-based hybrid approaches in classification and features selection. For incomplete information systems, rule generation by the GDT(General Distribution Table) approach and a default rule extracting method were proposed. Incremental algorithms are important, as data sources are increasingly in quantity. The group has investigated the …