Colloquium - Singh

The field of reinforcement learning (RL) has introduced into AI many abstract mathematical formulations of the AI problem, mostly by borrowing them from the fields of operations research and adaptive control. Has this been good for AI? I will argue that RL formulations have allowed us to ask and answer precise questions about many important AI issues as well as to bring a (mostly) principled design methodology to many application areas. I will illustrate these twin advantages through examples from my own work. In the first part of the talk I will provide an answer to a formulation of the "exploitation-exploration" tradeoff: should an agent exploit what it already knows or should it explore in the hope of learning something that leads to even greater long-term return? In the second part of the talk, I will describe why and how RL offers a powerful methodology for designing many human-computer interaction systems. I will illustrate this methodology through our design, construction and empirical valuation of NJFun (a system that provides telephonic access to a database of fun activities in NJ), and show that RL measurably improves NJFun's performance.

Hosted by Clayton Lewis.
Refreshments will be served immediately following the talk in ECOT 831.