Hell yeah!!! :mrgreen: :mrgreen: Finally a game which learns!! edit: question already answered in the vid :mrgreen:
I loved the detail they went into here, you guys rock. @sorian - any time you want to post more nerdy about how the AI learns, I know I'd love to read it! Thanks!
The AI learns alright... but not vs player. Imagine the ridiculous flow of data Sorian would have to deal with, not to mention the endless number of profiles.
You got it wrong. The neural net learns thats true, but it wont learn when playing against you or any other player. He will "teach" the neural net how to play before release (mainly by AI playing against AI doing random things until it figure out whats more and less effective). The AI we will run into in the end is set, it dosent evolve further on itself after release (They would have to patch it to improve it). (but it might change its tactics and such if units change in patches, since the neural net bases its decisions on the unit stats.)
Yes, it is exactly the way that things are presented in the video but they also said that may be the AI will learn during the party. As nothing is done for the moment, it can be or not...
I've always wondered how useful it would be for the AI to collate the data it makes while it plays against you, and then uploads that data to a grand central server to be merged into the main... Theoretically analyzing tactics that work well against human players and incorporating them into a better and better AI over time... Would be cool, but obviously the practical implementations may be a bit difficult
I looked into the possibility of having the AI learn in the field for Supreme Commander 2. In that case the problem boiled down to making sure all players had the same set of neural net data. Once things solidify more for the PA AI I will look at this option again.
I made a post that went over my GDC 2012 talk about neural networks on my blog here: http://soriandev.blogspot.com/2012/03/gdc-2012.html You can watch my GDC presentation here (no GDC vault membership required): http://www.gdcvault.com/play/1015667/Of ... n-Path-Non Just click "Neural Networks in Supreme Commander 2 - Mike Robbins" from the list on the left side (you will have to scroll down to find it). I also end up getting most of the questions during the Q&A session at the end of the session (just scroll down to the end).
just watched the gdc presentation. pretty interesting! perhaps this topic could be moved to the general discussion forum, since it may be interesting to non-backers to and also doesn't contain anything backer-exclusive information.
Well.. I do think the AI should learn a bit while the game is released. But more to prevent a situation where you can always beat the AI by doing a certain trick (eg a trick where the AI always makes a mistake). For example: you can always win against an AI using a certain plane, flying around his base and attack him from behind. The AI should recognize its mistake and take measures against it.
I remember there was a game called "Conflict Zone" and developped by a company (MASA) specialized in AI. The AI was based on neural network and it was learning and improving itself after each game. This way you could not use the same strategy during next game because if so, then AI was smart enough not to fall into the same trap. it's an old game and probably AI was not so sofisticated. However having the feeling that AI was learning from its mistakes was really fun.
Isn't it much easier with PA, as it has client-server model? I mean - you don't need to actually automatically share learning information between servers and there is no enemy-AI computations on client side, isn't it? (This will also lead to servers split into some kind of "PvE/PvP" groups, which is also good)