EpicPinkPenguin/procgen

Procgen Benchmark数据集包含16个程序生成的环境，每个环境的数据实例由观察、动作、奖励、完成和截断等字段组成。数据集分为训练集和测试集，分别包含900万和100万条数据。数据集是通过使用PPO强化学习算法在每个环境中训练25M步生成的。该数据集旨在衡量强化学习代理学习可泛化技能的速度，并强调实验的便利性、环境内和环境间的高多样性。

The Procgen Benchmark Dataset contains expert trajectories generated by a PPO reinforcement learning agent trained on 16 procedurally-generated gym environments. Each environment configuration includes features such as observation, action, reward, done, and truncated, with train and test splits. The dataset is designed for reinforcement learning tasks and includes performance metrics of the agent in each environment. The dataset creation process involves training the PPO agent for 25M steps in each environment, generating trajectories by sampling from the predicted action distribution. The Procgen Benchmark, developed by OpenAI, is designed to measure the generalizability of RL agents and is ideal for evaluating sample efficiency and generalization.