muzero-general VS open_spiel

I went looking for MuZero implementations in order to see how, exactly, they interact with the game space. Based on this one, which had the most stars in the muzero topic, it appears that it needs to be able to discern legal next steps from the current game state https://github.com/werner-duvaud/muzero-general/blob/master/...

So, I guess for the cards Forge has implemented one could MuZero it, but I believe it's a bit chicken and egg with a "free text" game like M:TG -- in order to train one would need to know legal steps for any random game state, but in order to have legal steps one would need to be able to read and interpret English rules and card text

You may want to use the trick described in https://arxiv.org/pdf/1805.11593.pdf as a Transformed Bellman Operator. Its efficiency is proved in MuZero original paper https://arxiv.org/pdf/1911.08265.pdf Appendix F. The implementation of that method you can find here: https://github.com/werner-duvaud/muzero-general Usage: muzero/models.py:649 (def support_to_scalar)

I have used this implementation from MuZero: https://github.com/werner-duvaud/muzero-general

+1 to taking a look at OpenSpiel. It has AlphaZero in C++ and Python, and there is even a PR open that allows running UCI (e.g. Stockfish) bot. You can also load chess via the OpenSpiel wrapper in muzero-general: https://github.com/werner-duvaud/muzero-general

When I looked up some community implementations, like that of Werner Duvaud on GitHub and Discord, hoping to make my own contributions to this effect, I soon found that I was hopelessly out of my depth as an amateur programmer, even with the help of some other sources like this walkthrough series. However, from what I could tell, most of the people working on this sort of thing seemed to be tackling relatively simple games. At first I thought this might be largely due to limitations in hobby time or computing power available to these users, but then I also noticed that, unless I have misunderstood something, it seems like the games are required to be rebuilt entirely in the engine of (this implementation of) MuZero, which would also obviously be a limit on the complexity of games chosen.

Windows support (Experimental / Workaround: Use the notebook in Google Colab)

There are several if you google "muzero github", e.g. https://github.com/werner-duvaud/muzero-general

Deep mind actually has a repository where they applied this algorithm for incomplete-knowledge games. You could use it for reference: https://github.com/deepmind/open_spiel/tree/master/open_spiel/python/algorithms

+1. You can also find an implementation of Clobber and AlphaZero (and many other basic RL algorithms) in OpenSpiel: https://github.com/deepmind/open_spiel

You can find a lot of code online if you look, and many of these applied to Poker. There's a general implementation of both in Python and C++ in OpenSpiel, with some examples applied to small poker games. It's nice code to learn from because the algorithms operate over generic game descriptions, so there aren't game-specific design choices mixed up with the implementation of the algorithms, and you can create your own poker game and just run them on it.

And many other additions and improvements. See all the details here: https://github.com/deepmind/open_spiel/releases/tag/v1.3

I would checkout the openspiel package. It's main focus is RL in games (multi-agent environments). You'll find RL examples there and games that are small enough to solve without deep RL. There's also a wide range of environments from fully cooperative to adversarial zero-sum.

Hi, you can check out OpenSpiel: https://github.com/deepmind/open_spiel/

as for algorithms , openspiel repository has few implementations some of these are not related to imperfect information games , and others are not for multiagent environment and others are tabular algorithms .

This includes single-agent Gymnasium wrappers for DM Control, DM Lab, Behavior Suite, Arcade Learning Environment, OpenAI Gym V21 & V26. Multi-agent PettingZoo wrappers support DM Control Soccer, OpenSpiel and Melting Pot. For more information, read the release notes here:

I've had some success using Action Masking - you can refer to here https://github.com/deepmind/open_spiel/blob/120420a74a69354d64c10b51cd129d4587f9f325/open_spiel/python/algorithms/dqn.py but for DQN you need to mask out q values for invalid actions (as well as masking them during prediction). In my case I'm able to place my mask in the observation so can fetch it quite easily during prediction but if that's not possible you could query it from the environment and store it in the replay buffer (like they do in the link I shared)

Take a look at this simple implementation: https://github.com/deepmind/open_spiel/blob/master/open_spiel/algorithms/minimax.cc