Neural net AI idea

I was thinking about the whole neural net thing which Sorian had been working with in SupCom2, how he had run the computer for a while so that the AI could learn what worked and what didn't. I wonder if this approach could be used with single player mode, where each time you play a game on a certain map, with a certain set of factions, the AI uses the outcome of the match to try and learn how to do better next time under those same conditions. So instead of using the neural net for AI vs AI battles, adapt it for human vs AI battles.
The number of battles with a given set of circumstances that are needed might be quite large, but perhaps AI vs human battles could be automatically uploaded as replays and analyzed by GPG central comps to determine what the best best AI vs human tactics are. Or if not have it centrally analyzed based on mutlpe different people, at least have it locally analyzed based on my battles vs the AI. At the very least, it would be interesting for me as a player to play vs the AI multiple times in a row with the same setup and see how the computer adapts after each match.
I don't know what I've suggested is novel or even feasible. Just wanted to know what some of you more knowledgable people think about it. If it could be properly impleneted, I think it would be a huge selling point for me and other like myself.|||When would it process all the information? If during the game, it impacts performance. And you dont keep the game running after youre done just to process the AI either. And if it considers old information that was stored during the game, it would both impact memory AND performance...|||I'm talking about processing it after the game, getting it ready for the next game. Don't know how long it would take to process it, though.
If I knew the computer was needed after the game to process the information, I personally would give it the time.|||Of course I dont know how the stuff works, but Im just saying that most people wouldnt leave their game on just for that. I also think that the kind of information the neural net uses right now - blob vs blob - couldnt be gathered ingame. Sorians neural nets just bash each other at enormously high game speeds for days. One little skirmish game is insignificant, Sorians PC processes that in a matter of minutes, maybe less. The only thing that a personal neural net would be good for would be analyzing your builds. Theres something to be gained here. But, food for thought: How would the multiplayer AI work then? Use the files from the host? Or just dont use that extra info?|||My idea for how the AI would work is that it only applies the new AI toward game setups in which it has previous experience, otherwise the stock AI would be used. Additionally, if you are playing a game against AI and are just dinking around, and not playing seriously, you can have the option to tell the comp to ignore that game for the purpose of developing a better AI.
Maybe after the game have the game ask if you want to upload the replay so that at GPG the learning AI could be fed this information and have more data points to look at. The central comp at GPG could be set up to continually process the new information and churn out AI updates on a semi-regular basis.
Or, alternatively, the AI could be set up so it only analyzes your specific games with it, and you always have the option to play with the stock AI instead of the adapting AI.
It would be cool to play an AI on one map, have the AI try one strategy to beat me, and if it loses that match, try an alternative strategy next game on that map until it can come up with a strategy in which it is most successful, or at least, does not lose as badly.|||I think only having these files locally is the way to go here, to adjust to YOUR strategy. Sending them to GPG would require them to have systems to analyze the entire game. Im sure Sorian would like that, but that would require a lot of extra work. Starcraft 2 has something like that, but instead of improving the AI, they use information in replays for balancing.|||Neph pretty much hit the problem on the head. This was brought up some time ago and sorians response was that then each AI would be different and when it came to multiplayer there would be issues involved.|||Nephylim|||They definately use the replays they automatically gather. The win-loss ratio of races and units being used in these games alone says enough. They also check every league to see if some units are perhaps too hard to use.|||My idea also involves reverting to the stock AI for multiplayer. Of course the AI of the two players that has been undergoing the learning process would be different, but the stock AI would not. The stock AI would always be the same among different players. So the stock AI is never erased, just a new AI that has been undergoing learning is added for the individual. If your learned AI ever gets screwed up somehow, you could always erase your learned AI, or select an option to use the stock AI and start again. If the whole learning AI is not working out for someone, perhaps they could check a box to keep it from ever learning and just keep using the stock AI.
The point here is that with enough options for the player, and if properly implented, it seems like it would be doable. I'm no programmer, but what I've outlined seems reasonable to me.|||It would be interesting if you had the method of automatically uploading the replays, and then having the computer analyse those later? But regardless, utterly impractical for GPG and the current AI is pretty good insofar as I am aware... (human snobbery inbound)|||Well, its better than the Starcraft 2 AI... which says something. Early game its easily defeated, but no AI has better macro than Sorian's|||This is probably not something that would be implemented in SupCom2, I am thinking more along the lines of future titles, such as King and Castles perhaps (if it ever gets made)|||Nephylim|||you forgot micro. The only thing the ai is not good at is where you have someone trained in build orders and 1v1 fighting it
you should put some build orders into the ai so it is good early game|||If only I had more time to go over build orders.|||It's an interesting thing that in order to get the optimal AI build orders... you'd have to wait until the metagame evolved in a new game.|||spuddyt|||Z32|||Ive noticed that the AI simply didnt learn anything about naval. When I spam just subs, the AI still spams lots and lots of cruisers. Almost like theyre destroyers, I'd say. I feel like most balance changes wont matter to the AI, because it doesnt use great unit mixes and micro anyways. Its all about the macro game.