ANDREW MOORE
Professor, Robotics and Computer Science
www

My research interests center on Machine Learning and Reinforcement Learning with a particular emphasis on learning algorithms for self-improving factories, robot motion planning, adaptive control, industrial engineering. My primary goal is to help make AI become sufficiently practical, powerful and robust that is routinely used in controls and manufacturing industries. My ongoing projects:

General Memory Based Learning. A key component of any autonomous system is that it is able to constantly monitor itself and notice unexpected behavior such as unmodeled dynamics, changing dynamics or, in some cases, equipment malfunction. To do this, we make the system learn from experience. Our research has produced a highly automatic learning mechanisms. The controller explicitly remembers all its sensory experiences (typically such data is in the form of large vectors of real-valued numbers) in a large database of cases. Novel search algorithms are used to persistently monitor the database to spot unexpected trends, relations between variables, and opportunities for controller enhancements.

Reinforcement Learning. This is an exciting new discipline in which systems learn to control themselves optimally based on arbitrary reward and punishment signals. The central issues concern processing such signals to develop optimal controllers. Our recent research involves scheduling search control during on-line planning to minimize wasted computations. Ongoing research extends this to the PartiGame algorithm, in which exploration and planning are scheduled in a multi-resolution manner. A recursive partitioning of state-space adapts itself in real time to yield fine detail in the critical regions, while remaining at a coarse granularity elsewhere. We are also investigating combining state-space triangulations and tree-search methods to combat to curse of dimensionality.

Auton: AI-based Evolutionary Operation. Response Surface Methods (RSMs) are a heavily used, statistical technique used in the optimization of expensive industrial processes. Such processes are typically characterized by: a set of controllable parameters, a noisy measurement of the result of running a process with the given inputs. We desire to find the set of parameters which optimizes some measure of the result. Experiments are expensive, but there is plenty of computation time available between each experiment. Current RSM practice is, for sensible reasons, a far from automatic process. It is very possible that techniques from Machine Learning and AI may be of use in RSM applications. We have embarked on the production of a minimal-human-supervision controller for such systems, which watches and adjusts process over extended periods of use: persistently designing safe experiments and whenever appropriate improving the process.

Dealing with Massive Datasets. In new research, we are examining the computational issues involved with massive datasets, including the Edinburgh-Durham Sky Catalogue of over a million galaxies. We are examining generalizations of kd-trees and R-trees for caching information sufficient to permit answers to extremely wide classes of statistical questions in near-constant time.