Embedded Files
Title: Non-Gradient Descent
Type: PhD Thesis
Author: Taranto, Aldo
Supervisors:
1. First: Dr. Bernardo Pereira Nunes
2. Second: A/Prof. Ron Addie
Institution of Origin: Australian National University
Qualification Name: Doctor of Philosophy
Number of Pages: TBC
Year: 2022 - Present
Digital Object Identifier (DOI): TBC
Abstract
A statement of the research problem can be expressed, simply as, can continuous ant colony optimization (ACO) be extended to form various enhanced ACO (EACO) algorithms? Can these EACOs then be used to survey n-dimensional nonconvex, non-smooth and dynamically changing loss landscape surfaces?
Two review papers were derived, one of them being the first survey paper on continuous ACO, and the other being the first survey paper on loss landscape surfaces that arise from machine learning (ML) algorithms. The methods and procedures utilized in this thesis examine the various artificial intelligence (AI) search methods of ACO and apply them to traverse the loss landscape surfaces. By using Ito diffusions, a Ito bridge stochastic differential equations (SDE) version of ACO was derived which we call IB-ACO. Together with use of KD-trees of algorithm theory, a hierarchical nest partitioning (HNP) version of ACO was derived which we call HNP-ACO. A third research paper was derived that takes these two EACOs together with some of the latest continuous ACO algorithms and has them traverse the loss landscapes. We see that by regulating the noise on the surfaces together with the local topology of various sized minima, we make some novel contributions to computational optimization and solvers for ML algorithms.
Finally, two research papers are derived that apply the EACOs, one to elevator route planning (ERP) and one to database search optimization (DSO), making a total of seven research papers. These applied papers together with the other theoretical papers constitute an original contribution to knowledge.
Keywords: ant colony optimization, loss landscape surface, manifolds, search, KD-trees
ANZSRC Field of Research 2020:
460202 - Autonomous agents and multiagent systems
460207 - Modelling and simulation
460210 - Satisfiability and optimisation
College: Engineering, Computing & Cybernetics
School: Computing
Page updated
Google Sites
Report abuse