Directly Forecasting Belief for Reinforcement Learning with Delays

This paper introduces a novel framework for directly forecasting beliefs in reinforcement learning scenarios with observation delays. By incorporating predictive capabilities into the belief state estimation process, the method achieves improved performance compared to traditional approaches that rely on historical information alone.

Authors: Qingyuan Wu, Yuhui Wang, Simon Sinong Zhan, Yixuan Wang, Chung-Wei Lin, Chen Lv, Qi Zhu, Jürgen Schmidhuber, Chao Huang (equal contribution)

@inproceedings{zhan2025directly, title={Directly Forecasting Belief for Reinforcement Learning with Delays}, author={Zhan, Sinong Simon and Wu, Qingyuan and Wang, Yixuan and Huang, Chao and Zhu, Qi}, booktitle={International Conference on Machine Learning (ICML)}, year={2025}, url={https://arxiv.org/abs/2505.00546} }