Planning from Pixels in Atari with Learned Symbolic Representations
This work provides a more principled and unsupervised method for learning features for width-based planning in Atari, which is significant for researchers working on reinforcement learning from high-dimensional observations.
This paper tackles the problem of planning from pixels in Atari games by learning symbolic representations using variational autoencoders (VAEs). The VAE-learned features, when used with RolloutIW, outperform the original RolloutIW and human professional play on Atari 2600, while significantly reducing the feature set size.
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, $π$-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.