Planning from Pixels in Atari with Learned Symbolic Representations

Andrea Dittadi, Frederik K. Drachmann, Thomas Bolander

Research output: Chapter in Book/Report/Conference proceedingArticle in proceedingsResearchpeer-review

42 Downloads (Pure)


Width-based planning methods have been shown to yield state-of-the-art performance in the Atari 2600 domain using pixel input. One successful approach, RolloutIW, represents states with the B-PROST boolean feature set. An augmented version of RolloutIW, pi-IW, shows that learned features can be competitive with handcrafted ones for width-based search. In this paper, we leverage variational autoencoders (VAEs) to learn features directly from pixels in a principled manner, and without supervision. The inference model of the trained VAEs extracts boolean features from pixels, and RolloutIW plans with these features. The resulting combination outperforms the original RolloutIW and human professional play on Atari 2600 and drastically reduces the size of the feature set.
Original languageEnglish
Title of host publicationProceedings of 35th AAAI Conference on Artificial Intelligence
PublisherAssociation for the Advancement of Artificial Intelligence
Publication date2021
ISBN (Print)978-1-57735-866-4
Publication statusPublished - 2021
Event35th AAAI Conference on Artificial Intelligence - Virtual Conference
Duration: 2 Feb 20219 Feb 2021
Conference number: 35


Conference35th AAAI Conference on Artificial Intelligence
LocationVirtual Conference


Dive into the research topics of 'Planning from Pixels in Atari with Learned Symbolic Representations'. Together they form a unique fingerprint.

Cite this