Planning from Pixels in Atari with Learned Symbolic Representations

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 B-PROST boolean feature set. An augmented version of pi-IW, shows that learned features can be competitive handcrafted ones for width-based search. In this paper, we leverage variational autoencoders (VAEs) learn directly from pixels a principled manner, and without supervision. The inference model trained VAEs extracts pixels, RolloutIW plans these features. resulting combination outperforms original human professional play on drastically reduces size