The Reinforcement Learning Competitions

Events such as RoboCup, the Netflix Prize, the Trading Agent Competition, the International Planning Competition, the General Game Playing Competition, the AAAI Computer Poker Competition, and the DARPA Grand Challenge have succeeded in raising awareness and stimulating research about their respective topics. The empirical, problem-oriented nature of these competitions can be an important counterweight to traditional research efforts, which often focus more on theoretical results and algorithmic innovation. Competition results provide a barometer for which approaches are popular, effective, and scalable to challenging problems. The competitive nature of the events provides an incentive to transform theoretical ideas into practical tools, enabling the field to reap the benefits of its scientific progress. The field of reinforcement learning (RL) (Kaelbling, Littman, and Moore 1996; Sutton and Barto 1998), is ripe for such a transformation. Broadly speaking, RL researchers aim to develop online algorithms for optimizing behavior in sequential decision problems (SDPs), wherein agents interact with typically unknown environments and seek behavior that maximizes their long-term reward. Many challenging and realistic domains can be cast in this framework (for example, robot control, game playing, and system optimization), so RL algorithms contribute to the broad goals of artificial intelligence. In recent years, many advances have been made in RL theory and algorithms, particularly in areas such as balancing exploration and exploitation, Articles