Generalized Weighted Path Consistency for Mastering Atari Games

Authors: Dengwei Zhao, Shikui Tu, Lei Xu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments are conducted on the Atari 100k benchmark with 26 games and GW-PCZero achieves 198% mean human performance, higher than the state-of-the-art Efficient Zero s 194%, while consuming only 25% of the computational resources consumed by Efficient Zero.
Researcher Affiliation Academia Dengwei Zhao Shanghai Jiao Tong University zdwccc@sjtu.edu.cn Shikui Tu Shanghai Jiao Tong University tushikui@sjtu.edu.cn Lei Xu Shanghai Jiao Tong University Guangdong Institute of Intelligence Science and Technology leixu@sjtu.edu.cn
Pseudocode Yes Algorithm 1: Sample Preparation for GW-PCZero; Algorithm 2: Weighted PC target t P C estimation
Open Source Code Yes 1The source code is available at https://github.com/CMACH508/GW_PCZero.
Open Datasets Yes Experiments are conducted on the Atari 100k benchmark with 26 games to evaluate GW-PCZero in diverse environments.
Dataset Splits No The paper refers to the 'Atari 100k benchmark' and '100k interaction steps' but does not specify explicit dataset splits (e.g., percentages or counts) for training, validation, or testing.
Hardware Specification Yes Experiments are conducted on 4 NVIDIA Tesla A100 GPUs with 16 CPU cores.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes We set cb = 1.0 and ca = 0.1 in Eq. (16). Totally 32 of different random seeds are used. Other hyperparameter settings are the same as Efficient Zero, as summarized in Appendix 5.