Machine Learning and Friends Lunch
Information-driven Inference in Resource-constrained Envronments
John Fisher
MIT
Abstract
Inference in resource-constrained environments is a challenging problem where optimal solutions are generally intractable. For example, Bayesian filtering in distributed sensor networks presents a fundamental trade-off between the value of information contained in a distributed set of measurements versus the resources expended to acquire them, fuse them into a model of uncertainty, and then transmit the resulting model. Approximate approaches have been proposed that treat a subset of these issues; however, the approaches are indirect and usually consider at most one or two future time steps. I will discuss a method which enables long time-horizon sensor planning in the context of object tracking with a distributed sensor network. The approach integrates the value of information discounted by resource expenditures over a rolling time horizon. Simulation results demonstrate that the resulting algorithm can provide similar estimation performance to that of the common "most informative sensor selection" method for a fraction of the resource expenditures.
Additionally, I will present performance guarantees bounding the
performance difference between optimal and approximate methods of
measurement selection for information-driven Bayesian filtering. Note
that the structure of these bounds applies to a variety of machine
learning problems including active learning and inference in graphical
models. Furthermore, it can be shown that the bounds are tight.
This is joint work with Jason Williams