Machine learning: inference under uncertainty
In order to make inferences, machines need to have manageable data. This has led Ali to a project involving the conversion of high-dimensional data to low dimensional data. Ali uses algorithms to constrain the high-dimensional data to a set of meaningful characteristics. He explains, “Images, text, speech, DNA arrays are all examples of high-dimension data. The volume of data is hard to analyze and hard to understand. The technique I’m using is to map the data in a low-dimensional way first and then look at it.”
The technique allows the analyst to keep the breadth of information, but distil it down to a depth that can be more easily clustered, classified or compared. This is important for machine learning because it helps machines learn to distinguish between sets of actions and make inferences about future actions.
Ali has designed algorithms for robotic navigation using the low-dimensional data. The environment and possible actions are converted to a low-dimensional map for the robot. He notes, “This technique gives the robot a subjective map of the real world. Actions are meaningful in this map and we can extrapolate the ultimate destination point for the robot based on what it has learned about the map and the consequence of actions .”
Ali has worked on the sequential decision problems with M. Bowling (Alberta) and D. Wilkinson (graduate student, UWaterloo). It has had success in robotics, but has possible applications in other in industries. “We’re looking for an opportunity to apply our algorithm and what we understand about high/low-dimensional data to a problem in health or finance,” says Ali.
University of Waterloo Mathematics, Annual Report 2005