Suggestion - Persistent ai "personalities"

So currently the ai is taught by playing off against itself over and over, learning what reactions are appropriate for certain situations (as far as I'm aware), but when the game is shipped they stop learning. Presumably the is because it would get too difficult but I have an idea for this (I know people are saying that the ai can't learn against humans but I don't see that the process would be much different (I may be wrong)).

Rather than the ai learning as a whole, I think it would be cool if you could go "set up new ai opponent" and chose a name, initial play style etc. but then that ai, we'll call him Bob, will learn and retain the memories of how he should try to beat you only from the games you've played against him. Bob would slowly get better and more difficult as you played him more. Then if he got too hard you could always set up a new one, or just play against the set ai.

Again, I don't actually know if this is possible but I think it'd be cool if it is.

So, from what I seem understand you're saying an adaptive AI that records matches you've done against it and attempts to counter noticeable strategies you've used, essentially making repeat strategies ineffective?

Isn't that the same technology used in IBM's chess playing supercomputer?

I would love this kind of AI for PA

I know a few games which have that kind of learning AI - and they get badass!

I still remember the awesome Parabot (Half Life bot) which learns from Players & other Bots...first few rounds are easy - because they need to learn EVERYTHING from you. That is shooting, walking, running, killing and more. After a few hours they adapt really brutal to you (it even get to places you would never imagine an AI would try to get - but a player would do) BUT if they encounter a new map/playstyle they need to learn and adapt like a real player.

I remember those UT 2004 bot tournaments of self written AI. People need to say which of them is Human and which not - surprise was that most AI behaved more human than AI xD

this could be a cool feature. how many battles can you win in a row against a learning ai

care to post a link to some videos?

I just want to point out that I didn't know regular gaming dudes could program a learning program - thought that was still in the bounds of crazy theories that kind of work.

Now I hope the robots dont make a skynet.....

They don't learn in the sense of human learning. Data enters their input (like, there are some bombers attacking my metal extractors) and they use their memory of the best outcome of previous times this has happened to choose the best response (Like sending interceptors). They don't actually have any awareness of what it is they're doing or why they're doing it.
So they are learning, but not intelligent.

I want the AIs to be based off the devs. That would be both awesome and hilarious.

Sorian has said that it would be possible to let players train the neural network, but it wouldn't make sense in multiplayer. If you train your ai to build a lot of one thing, another player in an ffa will have no problem finding a counter.

Yeah this would be purely for single player.

obligatory:
http://arstechnica.com/gaming/2011/...ermind-won-the-2010-starcraft-ai-competition/

and CoreWars: http://www.corewars.org/

I wish they allow us to modify or code own AI, like publishing some API for ai calls. Would be cool to code AI that owns Sorians AI. Something like core wars but as PA games.

I'm pretty sure Sorian is a fan of AI modding.

You never know, you do not know the depths of the hypocrisy that flourishes in the business right now. Many of the games coming out now that are completely locked down, encrypted, in the cloud etc. and otherwise utterly unmoddable in any way are coming from the people who got their jobs through modding.

brad ai. bots bots and more bots. mabe a tank? nah.

MOAR BOTS!!!

I would really like AI personalities. For example AI 1 is more into air whereas AI 2 is more into artillery. Or AI 1 is more into protective play, advancing inch by inch.. whereas AI 2 is more into hit and run strategy hitting you into the heard of your base.

I think Sorian has said this is going to be a thing down the line. As for a learning ai, I don't think neural nets are in yet.

They're in, but not on.

That i am also sure, but does he share that and let us fiddle with AI without game hacks?

And I am not buying games that are unmoddable, well i eventually buy them year or two later when they are in cheap bin. But I will not pay 50-60euro for game that can be finished in 2 evenings and will never have mods.

In fact, if I really like a game, I end up modding it a lot. Unless that is too hard or not at all possible of course... Which always disappoints me.