Razvan Nicolae Constantinescu
DSA-RL: Domain Space Adaptive with Reinforcement Learning for Univariate and Multivariate Timeseries Anomaly Detection.
Rel. Luca Cagliero. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
|
PDF (Tesi_di_laurea)
- Tesi
Accesso limitato a: Solo utenti staff fino al 2 Aprile 2028 (data di embargo). Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) |
Abstract
Time Series Anomaly Detection (TSAD) aims to identify anomalous points or intervals within long data series. Recent work on offline unsupervised TSAD largely follows a reconstruction paradigm in which a model encodes and reconstructs pointwise values and an anomaly score is estimated from the error. Time Series Language Models (TSLMs) bring broad pretraining knowledge to this task, but their training and inference can be prohibitively expensive. Lightweight architectures such as AutoEncoders (AEs) are cheaper, yet their performance depends on an anomaly-scoring function fixed a priori, which is a liability when anomaly-related information is scarce or the series contains anomalies of mixed types.
This thesis proposes DSA-RL, a cost-effective approach to offline unsupervised TSAD that couples an AE backbone with a Group Relative Policy Optimization (GRPO) reinforcement learning strategy
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
