Topic: LLM-backed AIs

We should certainly consider creating a fully neural network based AI player if we can. FTR the current AI also uses a very small neural net, but that requires way too much handholding from specialised code.

I'm less keen on human language LLM in the chat, both because I'm afraid that it would increase the hardware requirements too much, and because I doubt we can find one with clean, legally obtained training data with cleanly GPL compatible licence. In multiple languages.

Top

Quote

I also think that going all-in on machine learning for the computer players would be great to have. However even the current state involved lots of effort on the code side and lots and lots of contributor time to oversee the human-guided training sessions. We'd have to come up with a concept for a fully autonomous self-supervised AI training mode, in addition to the actual major changes to the 14000+ lines of existing AI code. So an absolutely worthwhile thing in my opinion, but which will also require enormous effort.

-1 for making AIs network clients, because we absolutely don't want to force players to connect to some server to play Widelands. Playing offline must never be penalised compared to playing online.
Also -1 for chat LLMs, because of 1) hardware requirements for an unnecessary piece of eye candy, 2) personally I find chatty computer players annoying, and 3) like all commonly available generative ML models the elsewhere discussed ethical and legal problems.

Top

Quote

Just my thoughts:

If we use a local AI for Widelands, we would/should not need a very large model. I.e. we could use a small model that can run in realtime on a typical CPU with a runtime performance not worse than the current computer players. That way we would be able to provide this locally to every user without raising our (currently very lenient) hardware requirements¹, which is the best possible outcome.

We could use pytorch or tensorflow as they also support CPU-based inference².
Using python for training is fine; our regression test suite runner is also written in python, so some interfaces to run Widelands programatically already exist.

We'll probably need to train a model from scratch just for Widelands though, unless there's a guaranteed-ethical pretrained model that would fit for our use-case.

If this works out, then in the long run it might ideally replace the existing computer players. Our computer player interface is modular so it is trivial to add new modes alongside the existing ones. The real work Widelands-side would be to define and provide an interface that the ML model can use: It will at least need to be able to query game state (map fields, map objects and their details, player statistics, global statistics, …), and have interfaces to send each type of player command, and should be automatically extensible so it won't panic when new features are added or the internal command syntax is changed.

Top

Quote