A COMPARISON BETWEEN BUCKET ELIMINATION AND MINI-BUCKET ELIMINATION PERFORMANCE
Keywords
No Thumbnail Available
Authors
Issue Date
2020-07-10
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
There are many different inference techniques for Bayesian networks. However, the intractability of Bayesian inference remains a problem for increasingly large networks. Approximate inference makes it possible to reduce the complexity, but guaranteeing a bound on the accuracy of this approximation is still intractable. So-called ‘anytime algorithms’ are a clever way to ensure the best result possible within any given time constraint. One such algorithm is the mini-bucket algorithm, which is an approximation technique based on bucket elimination. This thesis takes a look at a comparison between the mini-bucket algorithm and bucket elimination. When looking at the results of the tests, the mini-bucket algorithm performs well, even in the least accurate cases, and terminates significantly faster. Given the complexities of both algorithms and the results shown, the mini-bucket algorithm could in most cases completely replace the bucket elimination algorithm, since iteratively increasing the accuracy for the mini-buckets means the runtime of all previous runs, with less accuracy, becomes neglectable compared to the latest one. Compared to bucket elimination the increase in runtime is thus insignificant, but mini-buckets will always have an accurate approximation ready at all times during its runtime. Only when the network size is not an issue, bucket elimination will have the upper hand, as the conversion from bucket elimination to the mini-bucket algorithm is not necessary.
Keywords: Bayesian Networks, Inference, Bucket Elimination, Mini-Bucket
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen