Étude de la motivation intrinsèque en apprentissage par renforcement

Arthur Aubret 1 Laëtitia Matignon 1 Salima Hassas 1
1 SMA - Systèmes Multi-Agents
LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information
Abstract : Despite many existing works in reinforcement learning (RL) and the recent successes obtained by combining it with deep learning, RL is facing many challenges. Some of them, like the ability to abstract the action or the difficulty to conceive a reward function without expert knowledge, can be addressed by the use of intrinsic motivation. In this article, we provide a survey on the role of intrinsic motivation in RL and its different usages by detailing interests and limits of existing approaches. Our analysis suggests that mutual information is central to most of the work using intrinsic motivation in RL. The combination of deep RL and intrinsic motivation enables to learn more complicated and more generalisable behaviours than what enables standard RL.
Document type :
Conference papers
Complete list of metadatas

Cited literature [33 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-02272091
Contributor : Arthur Aubret <>
Submitted on : Wednesday, August 28, 2019 - 1:59:45 PM
Last modification on : Thursday, November 21, 2019 - 2:11:02 AM

File

JFPDA2019_2.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-02272091, version 1

Citation

Arthur Aubret, Laëtitia Matignon, Salima Hassas. Étude de la motivation intrinsèque en apprentissage par renforcement. Journées Francophones sur la Planification, la Décision et l'Apprentissage pour la conduite de systèmes, Jul 2019, Toulouse, France. ⟨hal-02272091⟩

Share

Metrics

Record views

87

Files downloads

92