Computer Science Colloquium
Technische Universität Berlin
Convergence Results in Reinforcement LearningThu 06.04.2006, 17:15, 60 minutes
AbstractThe term reinforcement learning refers to a category of machine learning techniques which deal with the estimation and maximization (through acting) of an abstract value named reward. The concept of reward driven learning is very general and hence it is possible to apply the techniques to a wide variety of problems: e.g. control systems for robots; adaptive websites, users provide feedback in terms of reward (very nice - very bad page) and an agent must learn to maximize this reward through generating the optimal website. This talk will be about the reward estimation problem with a focus on convergence speed analyzes of different methods. I will start with a short survey of basic concepts and terminology. After that I will discuss known convergence results and present new ones for two major techniques (temporal difference learning and monte carlo estimation).
Invited by Prof. Dr. Sepp Hochreiter
The Computer Science Colloquium is organized by the Department of Coputer Science at JKU, the Österreichische Gesellschaft für Informatik (ÖGI) and the Österreichische Computergesellschaft (OCG).