Hossein Favakeh
Study whether an LLM can improve the training of a small NN for indoor localization, compared to conventional optimization methods.
Rel. Mihai Teodor Lazarescu, Luciano Lavagno. Politecnico di Torino, Corso di laurea magistrale in Ict For Smart Societies (Ict Per La Società Del Futuro), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) | Preview |
Abstract
Abstract Knowing where a person is inside a building enables services such as assisted living, occupancy-aware heating and lighting, and access control. This thesis we investigated whether a large language model (LLM) can improve the training of a small neural network for indoor localization, and how its performance compares to conventional hyperparameter optimization methods. The quality of such a model depends on its hyperparameters, and each trial of manual tuning is expensive because every configuration must be trained from scratch. Moreover, whether a large language model (LLM) can act as a search strategy for tuning and improving conventional methods. A single-agent LLM optimizer is proposed: at each round it reads a compact, qualitative summary of the trials so far and proposes the next configuration, under strict validation that rejects any out of grid, or repeated proposal so that it can only reach configurations also reachable by the other methods.
It is benchmarked, under an identical protocol, the same dataset, neural network model, discrete search space, and trial budget, against four conventional methods: a fixed reference baseline, random search, a rule-based controller, and Optuna
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
