@article{Ayton_Asai_2022, title={Is Policy Learning Overrated?: Width-Based Planning and Active Learning for Atari}, volume={32}, url={https://ojs.aaai.org/index.php/ICAPS/article/view/19841}, DOI={10.1609/icaps.v32i1.19841}, abstractNote={Width-based planning has shown promising results on Atari 2600 games using pixel input, while using substantially fewer environment interactions than reinforcement learning. Recent width-based approaches have computed feature vectors for each screen using a hand designed feature set (Rollout-IW) or a variational autoencoder trained on game screens (VAE-IW), and prune screens that do not have novel features during the search. We propose Olive (Online-VAE-IW), which updates the VAE features online using active learning to maximize the utility of screens observed during planning. Experimental results across 55 Atari games demonstrate that it outperforms Rollout-IW by 42-to-11 and VAE-IW by 32-to-20. Moreover, Olive outperforms existing work based on policy-learning (π-IW, DQN) trained with 100 times the training budget by 30-to-22 and 31-to-17, and a state of the art data-efficient reinforcement learning (EfficientZero) trained with the same training budget and ran with 1.8 times the planning budget by 18-to-7 in the Atari 100k benchmark, without any policy learning. The source code and the appendix are available at github.com/ibm/atari-active-learning and arxiv.org/abs/2109.15310 .}, number={1}, journal={Proceedings of the International Conference on Automated Planning and Scheduling}, author={Ayton, Benjamin and Asai, Masataro}, year={2022}, month={Jun.}, pages={547-555} }