Methods for efficient, exact combinatorial computation in machine learning

Download Statistics

Downloads

Downloads per month over past year

Kayas, Ugur ORCID: https://orcid.org/0000-0003-3936-4090 (2023). Methods for efficient, exact combinatorial computation in machine learning. University of Birmingham. Ph.D.

Preview

Kayas2023PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (1MB) | Preview

Abstract

Combinatorial problems are common in machine learning, but they are often large-scale, exponential or factorial complexity optimization problems, for which exhaustive methods are impractical. Heuristics are typically used instead but these are not provably optimal, although they may produce a workable compromise solution in a reasonable time. On the other hand, dynamic programming (DP) is an efficient and broadly applicable tool that finds exact solutions to combinatorial problems. However, DP lacks systematicity as most algorithms are derived in an ad-hoc, problem-specific manner. In the literature, there are attempts to standardize DP algorithms, but they are either unnecessarily general (constructive algorithmics) or have limited applications to different problems (Emoto’s GTA).
In this thesis, we propose a rigorous algebraic approach that systematically solves DP problems either by deriving algorithms from existing ones, or by deriving them from simple functional recurrences. The main contribution is providing novel, exact solutions for combinatorial optimization problems in machine learning and artificial intelligence. Our novel formalism largely bypasses the need to invoke the often quite high level of abstraction present in classical constructive algorithmics, as well as providing algorithms that are provably correct and polymorphic over any semiring. These algorithms can be applied to any combinatorial problem expressible in terms of semirings as a consequence of polymorphism.
This approach also contributes to systematicity in embedding combinatorial constraints applying tupling to avoid the need for ad-hoc backtracking.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):