Monday, April 13, 1998
2:00 PM (refreshments 1:45)
RLE Conference Room, Room 36-428
EECS Special Seminar
Abstract
Sequential decision problems arise in a wide range of applications in computer science, econometrics, and engineering, where decisions are made serially as a set of data is revealed. In this talk, we motivate and introduce a general framework for such problems based on universal coding and machine learning and suggest a new paradigm for adaptive algorithms with a variety of applications in signal processing and communication.
In this context, we present an algorithm for linear prediction that is "twice universal" over parameters and model orders. For every data sequence, the sequentially accumulated prediction error is as small as that for any linear predictor, including one whose parameters are tuned to the data in advance. These results can be applied to a broad range of signal processing and communication problems, such as adaptive filtering or channel equalization.
This framework provides insight into the performance and robustness of some widely-used methods, such as the recursive least-squares algorithm. Additionally, the long-standing model order selection problem is mitigated by a shift in the philosophy used in adaptive algorithms. Rather than selecting or estimating the model, a weighted combination of all models can be constructed that achieves the best possible performance at the computational cost of a traditional algorithm.
|
Modified: Mar 19, 1998|
Current events|
Your comments
and inquiries are welcome.