Partially Observable Process Gym (POPGym)

Partially Observable Process Gym (POPGym) 是一个由剑桥大学开发的两部分库，包含15个部分可观测环境，每个环境有多个难度级别，以及13个记忆模型基线实现，是目前单个RL库中最多的。POPGym环境多样，产生较小的观察结果，使用较少的内存，并且通常在消费者级GPU上训练两小时内收敛。该数据集旨在解决RL在部分可观测环境中的挑战，提供了一个全面的基准，用于评估和比较不同的记忆模型在广泛任务上的表现。

Partially Observable Process Gym (POPGym) is a two-part library developed by the University of Cambridge, containing 15 partially observable environments (each with multiple difficulty levels) and 13 memory model baseline implementations — the largest count among standalone reinforcement learning (RL) libraries to date. POPGym features diverse environments, generates compact observations, boasts a low memory footprint, and typically converges within two hours of training on consumer-grade GPUs. This dataset aims to address the challenges of RL in partially observable environments, providing a comprehensive benchmark for evaluating and comparing the performance of different memory models across a broad range of tasks.