Unsupervised Learning of State Representation using Balanced View Spatial Deep InfoMax: Evaluation on Atari Games - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Unsupervised Learning of State Representation using Balanced View Spatial Deep InfoMax: Evaluation on Atari Games

Résumé

In this paper, we present an unsupervised state representation learning of spatio-temporally evolving sequences of autonomous agents' observations. Our method uses contrastive learning through mutual information (MI) maximization between a sample and the views derived through selection of pixels from the sample and other randomly selected negative samples. Our method employs balancing MI by finding the optimal ratios of positive-to-negative pixels in these derived (constructed) views. We performed several experiments and determined the optimal ratios of positive-to-negative signals to balance the MI between a given sample and the constructed views. The newly introduced method is named as Balanced View Spatial Deep InfoMax (BVS-DIM). We evaluated our method on Atari games and performed comparisons with the state-of-the-art unsupervised state representation learning baseline method. We show that our solution enables to successfully learn state representations from sparsely sampled or randomly shuffled observations. Our BVS-DIM method also marginally enhances the representation powers of encoders to capture high-level latent factors of the agents' observations when compared with the baseline method.
Fichier principal
Vignette du fichier
ICAART_2022.pdf (331.61 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03874522 , version 1 (28-11-2022)

Identifiants

  • HAL Id : hal-03874522 , version 1

Citer

Menore Tekeba Mengistu, Getachew Alemu, Pierre Chevaillier, Pierre de Loor. Unsupervised Learning of State Representation using Balanced View Spatial Deep InfoMax: Evaluation on Atari Games. International Conference on Agents and Artificial Intelligence (ICAART), Feb 2022, Porto, Portugal. pp.110-119. ⟨hal-03874522⟩
37 Consultations
25 Téléchargements

Partager

Gmail Facebook X LinkedIn More