Transfer Learning for Dynamical Systems Models via Autoencoders and GANs

Transfer learning has seen significant progress in the domains of supervised learning and reinforcement learning, yet there remains a noticeable gap in the area of regression. In supervised learning, various transfer learning methods have been developed to leverage knowledge from one task to improve...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Proceedings of the American Control Conference S. 8 - 14
Hauptverfasser: Damiani, Angelo, Lopez, Gustavo Viera, Manganini, Giorgio, Metelli, Alberto Maria, Restelli, Marcello
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: AACC 10.07.2024
Schlagworte:
Accuracy
Adaptation models
Autoencoders
Heuristic algorithms
Instance Transfer
Markov Decision Process
Model Transfer
Prediction algorithms
Reinforcement learning
Supervised learning
Transfer learning
ISSN:2378-5861
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Transfer learning has seen significant progress in the domains of supervised learning and reinforcement learning, yet there remains a noticeable gap in the area of regression. In supervised learning, various transfer learning methods have been developed to leverage knowledge from one task to improve performance on another, but these approaches are primarily designed for classification tasks. In the realm of reinforcement learning, many techniques focus on policy transfer, and those that do address samples transfer predominantly apply it within homogeneous contexts. This paper presents a novel algorithm that bridges the transfer learning gap between supervised learning and reinforcement learning while specifically addressing the regression problem of dynamical systems' model estimation. Our approach harnesses the feature extraction capabilities of autoencoders and the generative power of Generative Adversarial Networks (GANs) to train a mapping that facilitates the seamless transformation of samples between dynamical systems. This approach represents a significant advancement in the field of transfer learning, as it offers a versatile and effective solution for transferring regression models across heterogeneous domains. Our experimental results demonstrate the algorithm's efficacy and its potential to improve model generalization and adaptability in diverse scenarios with varying data distributions and dynamics.
ISSN:2378-5861
DOI:10.23919/ACC60939.2024.10644658