Inquiry About Creating New Game and Using it to Train Algorithm

[pyfb78]

I was wondering if there was any documentation available how I would create my own game or modify a game already present in the repository. Additionally, I was wondering how the algorithms present in the repository (specifically the ones that use reinforcement learning) were trained so that I could try to replicate some results in papers and then use the algorithms on modified games?

[lanctot]

Hi @pyfb78,

For the first question, check out the developer guide: https://github.com/google-deepmind/open_spiel/blob/master/docs/developer_guide.md. There are many examples under games/ and python/games and feel free to ask here. The best way to learn the core API is the tutorial video linked on the main page.

There are no trained agent in this repo, only algorithms. There is one reproducibility report that I know of but it does focus on the imperfect information games benchmarks: https://arxiv.org/abs/2103.00187

Depending on the complexity of your game, you may want to look into training frameworks that are intended for scale (RLLib, MAVA, MALib, etc. -- many of them support OpenSpiel games). The algorithms in OpenSpiel are not designed for scale, but mostly clean/easy to learn from demonstration for single machines.