Game Engine Rules (AI Training)
-
@beelee thanks! Gave it the old college try!
-
@Kindwind Too bad. I'm starting mine, I plays and completes games, but it still loses to pre built (fast) AI, it can beat easy. When you said that you beat the hard AI, do you mean that you were able to make a bot that was stronger Than the Hard AI that ships with Triple A? or something else.
If you can beat the Hard AI currently in game that's not a bad result at all.
-
@husky81 I was able to beat random not the hard ai. All kinds of odd bits and odd stalemates. The AI would buy all artillery it was really whacky. It was a good attempt on my part but I didn't know what I was doing math wise. Which in turn messed up my data. Finding signals is nigh impossible and a signal is something the ml really needs. The win conditions were another problem. Game state search was another problem. Aircraft carrier logic. Valid moves was another. The rust engine was a lot harder than it looked I was never really sure if mine was right. I worked on it for about a year or so.
-
@Kindwind I don't know what ai model you were using when you tried the project, but Opus 5.5 is better than what was available back then. I'm not sure if you had Fable back then. I don't have a Fable account, only Opus 5.5.
Opus was able to build a python rules engine pretty much in one shot (no GUI) so i can't see the games, but it will simulate games and write reports. My bot can play in Triple A as well by passing the moves generated by the python code to triple A and checking for rejected moves. That's how I validated my rules engine, and caught the errors and omissions around transports and aircraft carriers.
I'm not currently using machine learning, other than tuning some parameters my bot uses by running simulated games. But I may give a more ML approach a go in future.
-
On maps that are similar to the published A&A games, the Hard AI is only modestly better than the Fast AI. The Hard AI superiority is in games with more complex combat systems, where the simulation of the actual battles is a substantial advantage.
-
mplex co
Thank you for this note. I have seen that empirically. In testing the two in game ai's perform similarly. I was not fully aware of what made hard "better", but the ability to handle customized combat rules makes a lot of sense. I was aware that it simulated alot of battles.
-
@husky81 that is amazing! I had fable 5.1 and opus. What did was had Fable play some games and then export the games fable played to a triplea save file that I could open and then verify. I was able to look at the buys and the attacks and if transports were moving. Now there might be other ways that you know of that I don't. The project is a near PhD level project and if you can get it to work your name will go down in history as the 1st person to do so! That is how hard this is to do. It was really beyond me. I believe in you though! AI is getting good! After you get all the good data you can then move it over to a local LLM and that is where the good stuff is for our community! Peoples play would increase leaps and bounds. If you need help or want colab let me know I can lend a hand to help!
-
@husky81 Jev is really speaking to me on this as well but I don't know. Jev's set up seems much better over all suited to axis and allies. Just a thought!
-
speaking to me on this as well but I don't know. Jev
Jev seems a bit above my paygrade when it comes to coding.
Right now I have been hitting a bit of a wall. I have a bunch of systems that at least sound good on paper, but they have trouble functioning together. That is one issue that I'm having.
I have moved from the 1942 SE map to the minimap, which someone suggested. I played a game or two there i stomped the fast bot. I then asked Claude code to examine my behaviour and try to extract some rules to help it play better tactics. It does ok, and has scored a few wins vs fast. but it still either loses or fails to close most games.
-
A next idea that i want to try is having a sort of message board amongst my bot modules, so purchase, combat move, non combat move, ect. they can share requirements or requests with each other. Each module would run recursively on a loop and try to reach a consensus about the shape of the turn before doing anything.
There would have to be some type of executive module to harmonize everything to prevent infinite recursion. The bot would therefore formulate a plan for the turn before doing anything. After combat, non combat move and placement would have to run again, going off of the actual battle results and decide where the plan needs changing.
I'll ask Claude if he can build something like that the next time I get some more tokens.
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login