Stanford University, March 21-22, 1998
Reinforcement learning is the problem of improving an agent's behavior through trial-and-error interaction with an uncertain environment. Thus, reinforcement learning is more naturally characterized as a class of problems rather than a set of techniques.
One broad class of techniques involves learning a utility value function for states and/or state-action pairs the agent will encounter, as done in Q learning. Other approaches evaluate the policies directly, as in genetic and hill-climbing methods. One goal of this symposium is to foster greater communication between traditionally separate communities within the broader field of machine learning.
Another important objective of the meeting is to review, consolidate and stimulate new work in the application of reinforcement learning. This field has recently made important advances in theoretical understanding, particularly in the area of value-function techniques, and yet the number of real-world applications has remained relatively small.
However, there have been a number of notable successes, and these will provide an important focus for the symposium. In general, we need a better understanding of the characteristics of the problem domains that are well-suited to reinforcement learning methods. We intend the symposium will help resolve some of these issues, and lead to identifying others.
The symposium will take place on Saturday, March 21, and on Sunday, March 22, just before the AAAI Spring Symposium, at Stanford University in Room 107, History Corner (Building 200).
Attendance will be by invitation only, but there is no registration fee. There will be 15 invited speakers presenting at the symposium over two days. We anticipate that a similar number of non-presenting attendees will participate in the meeting.