Gymnasium VS popgym

"Show HN: Ghidra Plays Mario" (2023) https://news.ycombinator.com/item?id=37475761 :

[RL, MuZero reduxxxx ]

> Farama-Foundation/Gymnasium is a fork of OpenAI/gym and it has support for additional Environments like MuJoCo: https://github.com/Farama-Foundation/Gymnasium#environments

> Farama-Foundatiom/MO-Gymnasiun: "Multi-objective Gymnasium environments for reinforcement learning": https://github.com/Farama-Foundation/MO-Gymnasium

https://github.com/Farama-Foundation/Gymnasium#environments

Farama-Foundatiom/MO-Gymnasiun:

https://github.com/Farama-Foundation/Gymnasium - A framework Python library to build and train your own AI to play games

The only change in the environment from the one here is the reward function which is given its return value using the following code snippet (replacing lines 648-672 in the above url):

This may be useful: https://github.com/Farama-Foundation/Gymnasium

Here are all the environments. So for example, if instead of Hopper-v2 you want the acrobat environment from classic control you can write: env = gym.make('Acrobot-v1')

This release also includes a large number of documentation updates, minor bug fixes, and other minor improvements; the full release notes are available here if you’d like to learn more: https://github.com/Farama-Foundation/Gymnasium/releases/tag/v0.28.0.

Thanks! It really depends on the task that you want to implement. But in general, sticking to the standard gymnasium API is important. If you want to implement a 2D environment then PyGame is promising. If it's more like a game, check out Unity ML-Agents or Godot RL Agents. Anything simpler can also be just pure python code. You also need to carefully design your observation space, action space and reward function. My advice is to explore design choices of related environments.

You can read the release notes here: https://github.com/Farama-Foundation/Gymnasium/releases/tag/v0.27.0. You can upgrade from 0.26 without any changes unless you're doing something very uncommon; this is how releases will generally be going forward.

POPGym is based on RLlib and has two linear transformers and five or six RNN variants, including LSTM. I've found that transformers tend to perform pretty poorly in RL when compared to RNNs.

Code: https://github.com/proroklab/popgym

Have you seen this other ICLR paper, POPGym? Paper: https://openreview.net/forum?id=chDrutUTs0K Code: https://github.com/smorad/popgym

https://github.com/smorad/popgym contains 15 partially observable gym environments, but they use discrete actino spaces. I've verified that memoryless models (e.g. PPO+MLP) cannot solve these tasks, except for the navigation ones.