System Identification and Algorithms

A Description of the Collaborative Research Program's System Identification and Algorithms Project Cluster goes here.

Project 1: Object Detection

Project 2: On-line Index Selection

Project 3: Recursive Parameter Estimation Algorithm

Project 1

Location-based Object Detection

Members: Damian Eads, David Helmbold, James Theiler, Caroline Connor, Gary Grider

Object detection, a popular problem in computer vision, is often solved indirectly with pixel or sliding window classifiers. We have developed a new learning algorithm for learning and predicting directly in (x, y) location space, a more natural domain for some kinds of object detection. A new framework for boosting is devised: it combines a weighted ensemble of weak location predictors into a master predictor. By minimizing a spatially-oriented loss function, we directly optimize in the location problem domain. We introduce Hit-or-Shift (HoS) filter, an abstraction that takes any confidence-rated object detector as input, and outputs structured predictions that can easily be combined into a location ensemble. Algorithms are introduced to find HoS parameters that provably minimise a bound on a spatial loss function

Project 2

On-line Index Selection for Physical Database Tuning

Members: Karl Schnaitter, Alkis Polyzotis, John Bent

Index selection is a crucial component of performance tuning for a relational database. Traditional approaches to index selection use a "representative workload" that describes the queries that are expected in the future, and attempts to optimize the indexes for the this workload. It can be difficult for a database user to provide such a representative workload in advance, and the workload may evolve over time. Thus, we propose an on-line approach to index selection that can adapt to changes in the workload. We describe a "semi-automatic" algorithm that observes the incoming queries and makes index recommendations for the database administrator, who may choose to implement the recommendations or not. The administrator may also provide feedback on the recommendations, and the algorithm takes these preferences into account.

Project 3

Recursive Parameter Estimation Algorithm

Members: Janelle Yong, Don Wiberg, John Galbraith

Project summary to be provided.