We design novel algorithms to solve very large scale data mining problems from rich and diverse information sources. Our active learning and experiment design techniques allow us to maximize the benefit of expert labeling while minimizing the time and cost of this effort. We apply this research very broadly: applications include terrain classification from aerial multi-spectral imagery, learning based obstacle detection for robotic applications in challenging terrain and modeling and prediction of skid-steer vehicle motion.

We focus particularly on machine learning for problems of decision mining, reinforcement learning, and learning control. In these applications, the learning algorithm's decision has an immediate effect on the environment it interacts with. Our algorithms have been applied to the optimal control of robotic manipulators, learning-based obstacle avoidance, route-planning and re-planning, and automated drug discovery.

•   Active learning and experiment design •   Supervised, unsupervised and reinforcement learning

•   Learning for control •   Decision mining

Automated identification of urban imagery

Neural-net based optimization for bucket operation

Human ID in thick vegetation