Order Matters: Agent-by-agent Policy Optimization

Authors: Xihuai Wang, Zheng Tian, Ziyu Wan, Ying Wen, Jun Wang, Weinan Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate A2PO, we conduct a comprehensive empirical study on four benchmarks: Star Craft II, Multiagent Mu Jo Co, Multi-agent Particle Environment, and Google Research Football full game scenarios. A2PO consistently outperforms strong baselines.
Researcher Affiliation Collaboration Xihuai Wang1,2 , Zheng Tian3 , Ziyu Wan1,2, Ying Wen1, Jun Wang2,4, Weinan Zhang1 1 Shanghai Jiao Tong University, 2 Digital Brain Lab, 3 Shanghai Tech University, 4 University College London
Pseudocode Yes Algorithm 1: Agent-by-agent Policy Optimization (A2PO) and Algorithm 2: Agent-by-agent Policy Optimization (Parameter Sharing)
Open Source Code Yes The source code of this paper is available at https:// anonymous.4open.science/r/A2PO.
Open Datasets Yes Star Craft II Multi-agent Challenge (SMAC) (Samvelyan et al., 2019), Multi-agent Mu Jo Co (MA-Mu Jo Co) (de Witt et al., 2020), Multi-agent Particle Environment (MPE) (Lowe et al., 2017)3, and more challenging Google Research Football full-game scenarios (Kurach et al., 2020).
Dataset Splits No The paper evaluates reinforcement learning agents in simulated environments (Star Craft II, Mu Jo Co, MPE, GRF). These environments involve continuous interaction and episodic learning, not static datasets with pre-defined train/validation/test splits. Performance metrics are gathered from agent-environment interactions rather than partitioned datasets.
Hardware Specification No No explicit hardware specifications (e.g., specific GPU/CPU models, memory details) are mentioned for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, PyTorch version, CUDA version).
Experiment Setup Yes We tune several hyper-parameters in all the benchmarks, other hyper-parameters refer to the settings used in MAPPO. cϵ are selected to be 0.5 in all the tasks. (B.4 Hyper-parameters) And subsequent tables like Table 7, 8, 9, 10, 11 detail specific hyperparameter values for each task (e.g., ppo epoch, actor lr, critic lr, λ, ϵ).