Distributional Reinforcement Learning with Regularized Wasserstein Loss

Authors: Ke Sun, Yingnan Zhao, Wulong Liu, Bei Jiang, Linglong Kong

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show that Sinkhorn DRL consistently outperforms or matches existing algorithms on the Atari games suite and particularly stands out in the multi-dimensional reward setting.
Researcher Affiliation Collaboration 1University of Alberta, Edmonton, Canada 2 Harbin Engineering University, China 3Huawei Noah s Ark Lab
Pseudocode Yes Algorithm 1 Generic Sinkhorn distributional RL Update; Algorithm 2 Sinkhorn Iterations to Approximate Wc,ε; Algorithm 3 Sinkhorn Distributional RL
Open Source Code Yes Code is available in https://github.com/datake/Sinkhorn Dist RL.
Open Datasets Yes We substantiate the effectiveness of Sinkhorn DRL as described in Algorithm 1 on the entire 55 Atari 2600 games.
Dataset Splits No The paper states that algorithms are evaluated over 40M training frames and results are averaged over three seeds, but it does not specify explicit train/validation/test dataset splits.
Hardware Specification Yes We run our experiments on multiple NVIDIA 3090 Ti GPUs
Software Dependencies No The paper mentions building algorithms based on a 'well-accepted Py Torch implementation' and re-implementing MMD-DQN based on its 'original Tensor Flow implementation', but it does not specify version numbers for these or other key software dependencies.
Experiment Setup Yes For a fair comparison with QR-DQN, C51, and MMD-DQN, we use the same hyperparameters: the number of generated samples N = 200, Adam optimizer with lr = 0.00005, ϵAdam = 0.01/32. In Sinkhorn DRL, we choose the number of Sinkhorn iterations L = 10 and smoothing hyperparameter ε = 10.0 in Section 5.1 after conducting sensitivity analysis in Section 5.2. Guided by the contraction guarantee analyzed in Theorem 1, we use the unrectified kernel, specifically setting c = kα and choosing α = 2. We evaluate all algorithms on 55 Atari games, averaging results over three seeds.