HAPPO HATRPO

HAPPO HATRPO is a reinforcement learning resource focused on scalable deployment in multi-agent environments. The project provides trained model checkpoints and code for research involving the Inter-Class Actor-Critic approach, enabling changes to the agent count at deployment without requiring retraining. It is designed to support research snapshots and experimentation with scalable multi-agent reinforcement learning. The available files include actor and critic models for each agent class, specifically using a count-invariant graph neural network (GNN) with attention mechanisms for the actor, and a GNN class-factored central critic for training. Optimizer states are also provided to facilitate resuming training, along with configuration metadata detailing aspects such as map, classes, and hidden size. The tool targets scenarios involving the StarCraft Multi-Agent Challenge (SMAC), with provided checkpoints for maps like 2s3z and 3s5z, which feature different team sizes but the same unit classes. Additionally, results files such as zero-shot transfer win rates are included, supporting research into scaling and transferability in multi-agent systems. HAPPO HATRPO is implemented using PyTorch and supports integration with TensorBoard for monitoring training metrics. The project is released under the MIT License, allowing open access and modification. It is delivered through downloadable files and code, with further resources accessible via its GitHub repository. The tool is intended for researchers and practitioners working in multi-agent reinforcement learning, particularly those interested in scalable deployment and experimentation with agent class variations in environments like StarCraft. No information is provided about commercial pricing or proprietary licensing, as the evidence indicates the use of the MIT License. The project is positioned within the class of reinforcement learning tools, with a particular emphasis on scalable, multi-agent deployments and research reproducibility.