Adaptive/learning AI

I will be playing against the AI a lot so I have an interest to see a good AI. But with all games the AI is pretty static, once you understand it the games begin to be easy. Often more than one AI is "created" and a random AI is selected before the game starts so there is an element of surprise.

But it would be rather nice to have an AI that evaluates your gameplay after a match and changes his play style for the next match. The effect will be an easy AI in the beginning which will "grow" with every match you play against.

I understand the work that’s involved but in order to give the modding community the option to make such functionality I would like to see the option to save a match which can be used to evaluate and the AI modable so the modding community can make a program that can evaluate the match data and generate a new AI based on the played match.

OHGOD. Don't give Sorian ideas.

That is basically what a neural net AI does, which I think it'll be.
SupCom2 had a neural net AI, although I believe it stopped learning when they shipped the game.

It still knows enough to totally screw you however, bloody OP planes...

I definitely agree we need some mechanism of making the AI less predictable. And to have one which learns from its games would make every match against an AI different and exciting - almost like playing against a human. Indeed, it would probably make an AI harder too.
Not to mention that if the AI could truly teach itself, then once the system was set up it could potentially need less formal programming. I think that good AI will be one of the hardest things to get right about this game, and one of the most important.

This might be difficult to do, since the game will be moddable, there should also be support for Custom Units in the adaptive AI. So much for Sorian to do. Poor guy XD

Sorian should have asked what we wanted BEFORE he accepted the new job :mrgreen:

I would settle for one that could adjust it's playstyle mid-game against what you use against it.

I only say this because they will probably update the AI regularly against in-game balance and player strategy.

Yeah, as long as they build in the AI with a multitude of strategy's the players use, it should be great.

Hm, learning server-based AI? It makes sense, but it will separate servers network into
servers with good AI (where many people-against-AI battles are played) and servers with bad AI (opposite). And with more learning with will became only worse. Maybe it makes sense, but community separation is not good thing.

And, taking some people attitude into account, I don't believe it would be high rate of study data exchange. So, maybe, it should be requirement to share such data between servers in same lobby.

Would be good if the AI is learning himself from other individuals.
Like the robot from : The Incredibles
But with more intelligence overhaul

Well, since it is a client/server approach, it may well be possible for the AI to learn from everyone who plays online! Once a game has finished, another server could analyse the game—focus on select battles, or simply record army composition, who won and by how much—and the AI could replicate that in later games (responding to scouting information, etc.).

What do you think Sorian? An AI that keeps up with the so-called "meta game".

Has no one thought...
Adapitve/Learning AI? In control of multipul planets and 1000's of death-bots?
Dose anybody see a Skynet coming through?

No, I see an Ubernet.

I guess that's why they called it Ubernet.

and i want world peace and a cure for cancer.

Well, some people's goals are more easily achieved than other's.

if i understand it properly(which i doubt i do), youere talking about the basics of a nerual net, which is how an AI learns to Be an effective AI, sorian used one when creating the supcom AI, because the game will be played through servers i suppose it could keep analysing and learning on the ubernet servers and become truly ferocious

Thats not how an neural network works.

A neural network is able to learn to make better DECISIONS from a given (fixed) set of options, based on environmental parameters. The "AI" has no awareness of what those options are or what the parameters mean, it just "learned" what option leads to the best result in which situation.

However, it won't ever be able to learn "new moves". Thats not neural networks, thats genetic algorithms. It just won't be able to "watch" human players and mimic their gameplay or to develop a new strategy, it will only be able to choose the most appropriate strategy from a fixed set based on experience.

So when talking about an adaptive AI, we are talking about an AI which has several strategies hard-coded and is capable of switching between those strategies. You can teach the AI to switch more "intelligent" between strategies, but you can't teach the AI whole new strategical concepts.


Creating an real AI based on genetic algorithms is far of the limits, i highly doubt that there is anyone at Uber how has the required knowledge (or experience) to do such thing.

And: It's not something you wanna do on a live server. A genetic algorithm requires thousands of iterations before it produces an usable result and even then it won't become "stable". Most times the AI will screw up royally because it tries new approaches, only very few will show to be successful. The only efficient way to "train" a genetic AI is to let it compete with other genetic AIs.

There remains only one huge drawback: An AI produced this way may show "interesting" behaviours, but it is neither efficient (the "evolution" leaves many inefficient leftovers) nor will it be able to compete with real players who have access to a brain which had millions of years to evolve into one extremely fast adapting machine. Modern computers don't have the necessary computational power to compete with that (there are several million synapses participating in every decision, a PC can only simulate a few dozens to hundreds).

Learning is fine as long as the AI doesn't become perfect at the game and has no flaw in how it plays, being the AI's advantages of perfect micro management they shouldnt be perfect at the game.

Oh, micromanagement is something the AI should and must be perfect at

If you want to abuse flawed game mechanics by micro management in the intent to gain an unfair / unintended advantage (like making units virtually immortal by abusing range system or dodging missiles), well, thats not the way a strategy game is meant to be played, it could even be considered bugusing. Since you can't solve this problem by simplifying the mechanics without making the game unattractive, you have to solve this problem by balancing the game while using "perfect" behaviors on both sides, and that requires the support of an flawless(!) unit AI.

Just saying in case you still haven't understood what all the ranting about "more intelligent units" is about.


And when talking about the strategic AI: Don't worry, you will never see a flawless AI at that level, not within the next few years. The game is so much more complex then chess, and we still haven't gotten a perfect (as in unbeatable when starting with a certain color) chess AI yet, despite over 30 years of research.