Stumbled on this article in the wiki while looking for a copy of the User Guide. Was wondering how up to date everyone considers this to be. I’ve been experimenting with having AI synthesize board game moves. Right now I’ve managed to get ChatGPT to install VASSAL in its environment and play a full game of “Napoleon at Waterloo” against itself. Having it attempt to use the GUI was glitchy but it can write valid .vlogs as a workaround. This can cause unexpected formatting in the chat log but nothing major. I did run into bugs when I asked it to flop down combat arrow markers to help me track combat.
Wondering if anyone else has gotten up to similar mischief.
Does Vassal support computer opponents? - Documentation - Vassal
I’ve been looking into this some but the current Vassal does not have the best hooks for the game state and map contents. The 4.0 descriptions explain how that version will have a much better API.
Having the model play directly seems expensive for anything of any size since the decision space is huge. I know there a specialized models that do well for some games and these can be coupled to an LLM for additional capabilities (see all the Diplomacy research over the last 5 years). So having an AI help build an engine (better with a learning engine) with an LLM tie in for higher level considerations seems like it should be the best approach. The core design on the game play engine side needs some real thought to make all this work well enough to create at least a moderately interesting opponent.
Obviously it’s not as good as dedicated functionality like an API, but having the LLM run VASSAL and simply confect .vlogs directly has been working for me.
Re: cost, I have the cheapest paid ChatGPT plan and have not triggered any plan limits/timeouts/etc. I have had it run around 150 full games of NaW and my LLM usage analytics haven’t budged. It just seems like writing .vlogs is very cheap.
Quality of play is not great but I’m OK with that. I’m treating this like teaching a new player or an assisted solitaire game. For now I have given it scans of AH’s The General and asked it to attempt to generalize from that, knowing that the corpus would largely lack material related to NaW. I’m going to give it scans of Moves at some point and see if that turns it into a shark but for now I’m interested in how well it can generalize.
The complexity of the decision space goes like a factorial or power or product of the game elements compounded by the depth of the rules. So, Napoleon at Waterloo is reasonable since it has what, 200 hexes and 30-ish units with very simple rules. Much bigger and I don’t think you can handle generically at reasonable cost with a direct LLM approach.
Interesting investigation to see where the breakpoints are. .
Your point here is one that I see made repeatedly. I get it, but…my sense is that humans grab the rulebook and etc. and synthesize a move from that textual soup so the LLM may be able to also do so. Certainly humans are not conducting a full analysis of every possible board state and every possible action for each board state, right? They’re doing something else.
The thing that’s been interesting is that I picked NaW for basically the reasons you state. However, it’s proven more complex than I’d reckoned because of the ZOC-trapping and such that’s part and parcel of these old games.