Robust Approximate Reinforcement Learning
Speaker: John Langford , Toyota Technological Institute at Chicago
Date: October 14 2003
In the "real world" the space of an agents' states or observations is not practically enumerable, implying that exact algorithms to optimize an agents expected reward are impractical.
One common solution is to derive approximate forms of the exact algorithms, such as approximate policy iteration. Unfortunately, these approaches tend to be nonrobust since the sample complexity (or number of interactions with the world) required to guarantee success remains proportional to the size of the state space.
Another solution (currently being pursued by no less than 10 people) is to reduce reinforcement learning to classification, for which many algorithms and performance guarantees are _not_ explicitly or even implicitly dependent on the size of a state space. I will discuss how to do this, what theoretical guarantees can be transferred from classification, and show some empirical results suggesting this approach works well in practice.
See other events that are part of AI Colloquium Series Fall 2003
See other events happening in October 2003