Understanding non-convex optimization problems and stochastic optimization algorithms

This thesis presents significant contributions in the field of iterative stochastic heuristics. Various aspects related to the comparison and improvement of optimisation algorithms are addressed. Firstly, a methodology is proposed to fairly compare the performance of algorithms run on different mach...

Descripción completa

Detalles Bibliográficos
Autor: Arza, E.
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2023
País:España
Institución:Basque Center for Applied Mathematics (BCAM)
Repositorio:BIRD. BCAM's Institutional Repository Data
OAI Identifier:oai:bird.bcamath.org:20.500.11824/1836
Acceso en línea:http://hdl.handle.net/20.500.11824/1836
http://hdl.handle.net/10810/66155
Access Level:acceso abierto
Palabra clave:artificial intelligence
informatics
Descripción
Sumario:This thesis presents significant contributions in the field of iterative stochastic heuristics. Various aspects related to the comparison and improvement of optimisation algorithms are addressed. Firstly, a methodology is proposed to fairly compare the performance of algorithms run on different machines, ensuring an equitable allocation of computational resources. Additionally, a methodology based on stochastic dominance is introduced to compare the performance of optimisation algorithms as random variables. Furthermore, the relationship between Hamming distance and the quadratic assignment problem is analysed. A general early stopping method for learning policies in episodic problems, which does not require specific problem information, is developed. In summary, this thesis contributes to the understanding and improvement of iterative stochastic heuristics in the field of optimisation.