Mask Atari

Mask Atari是一个基于Atari 2600游戏的新型基准数据集，旨在通过深度强化学习解决部分可观测马尔可夫决策过程（POMDP）问题。该数据集通过引入可控、可移动和可学习的掩码作为目标代理的观察区域，特别在POMDP中的主动信息收集（AIG）设置中，构建了一个模拟环境。Mask Atari提供了一个挑战性的、高效的基准，用于评估专注于上述问题的各种方法。此外，掩码操作尝试将人类视觉系统的感受野引入代理的模拟环境中，这意味着评估不会因感知能力而产生偏差，而是纯粹关注方法的认知性能，与人类基准进行比较。该数据集适用于复杂、目标导向、视觉丰富的任务，旨在解决POMDP环境中的AIG问题。

Mask Atari is a novel benchmark dataset based on Atari 2600 games, aimed at addressing partially observable Markov decision process (POMDP) problems via deep reinforcement learning. This dataset introduces controllable, movable, and learnable masks as the observation regions for target agents, and constructs a simulated environment specifically for the active information gathering (AIG) setting in POMDPs. Mask Atari provides a challenging and efficient benchmark for evaluating various methods focused on the aforementioned problems. Furthermore, the masking operation seeks to introduce the receptive field of the human visual system into the agent's simulated environment, ensuring that evaluations are not biased by perceptual capabilities and purely focus on the cognitive performance of the method, enabling direct comparison with human benchmarks. This dataset is suitable for complex, goal-oriented, visually rich tasks, and aims to solve the AIG problem in POMDP environments.