Multi-task learning with modular reinforcement learning - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Multi-task learning with modular reinforcement learning

Jianyong Xue
Frédéric Alexandre

Résumé

The ability to learn compositional strategies in multi-task learning and to exert them appropriately is crucial to the development of artificial intelligence. However, there exist several challenges: (i) how to maintain the independence of modules in learning their own sub-tasks; (ii) how to avoid performance degradation in situations where modules' reward scales are incompatible; (iii) how to find the optimal composite policy for the entire set of tasks. In this paper, we introduce a Modular Reinforcement Learning (MRL) framework that coordinates the competition and the cooperation between separate modules. A selective update mechanism enables the learning system to align incomparable reward scales in different modules. Furthermore, the learning system follows a "joint policy" to calculate actions' preferences combined with their responsibility for the current task. We evaluate the effectiveness of our approach on a classic food-gathering and predator-avoidance task. Results show that our approach has better performance than previous MRL methods in learning separate strategies for sub-tasks, is robust to modules with incomparable reward scales, and maintains the independence of the learning in each module.
Fichier principal
Vignette du fichier
Multi_task_learning_with_modular_reinforcement_learning_.pdf (547.38 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03718157 , version 1 (08-07-2022)

Licence

Paternité

Identifiants

  • HAL Id : hal-03718157 , version 1

Citer

Jianyong Xue, Frédéric Alexandre. Multi-task learning with modular reinforcement learning. SAB 2022 - 16th International Conference on the Simulation of Adaptive Behavior, Sep 2022, Cergy-Pontoise / Virtual, France. ⟨hal-03718157⟩

Collections

CNRS INRIA INRIA2
67 Consultations
131 Téléchargements

Partager

Gmail Facebook X LinkedIn More