Adversarial MDPs

该数据集包含了一系列的剧集，这些剧集中的任务选自一组4个有限视野的MDP（马尔可夫决策过程）。每个MDP在一个4x4的网格上表示，具有特定的奖励结构。每个剧集都具有特定的结构，奖励只存在于网格的角落位置。初始状态是固定的，该数据集旨在通过预期累积奖励来评估算法。其规模包括4个MDP，200个剧集，以及16个状态。任务类型为对抗性在线多任务强化学习。

This dataset consists of a series of episodes, where the tasks in each episode are selected from a set of 4 finite-horizon Markov Decision Processes (MDPs). Each MDP is represented on a 4x4 grid and features a specific reward structure. Each episode follows a defined structure, with rewards only present at the four corner positions of the grid. The initial state is fixed across all episodes. This dataset is designed to evaluate algorithms based on expected cumulative rewards. It includes 4 MDPs, 200 episodes, and 16 states total. The task type is adversarial online multi-task reinforcement learning.