I am quite familiar with traditional Machine Learning and Deep Learning (especially Computer Vision), I have been mainly using TensorFlow2/Keras and some Pytorch.
Recently, I am learning Deep Reinforcement Learning, after attempting to write some of the algorithms with TensorFlow2 + pure Python, I heard that there are higher-level libraries for DRL like TF-Agent and Dopamine. However, it seems like there are way too many of them and I cannot figure out which is best for me, considering each of them is better at certain fields/aspects. I am looking for a Python library that is mainly good at (by good at, I mean it provides sufficient features/functions/algorithms, relatively efficient implementations, and people are more likely to use it for the listed purposes):
Video Gameplays (for example, playing Atari games, I've heard that some are better at research, some are better at robotics, so that's why I have this preference listed)
TensorFlow2-based preferred (so I can build models or use functions from TensorFlow2)
Support Deep Learning Model running on GPU
Support Not just Simulations like Gym, but actual computer games
Have decent documentations
Any help and recommendations would be great! Thanks.