I searched the forum, and watched the Kickstarter/Livestream videos again, but didn't find an answer to my question. Does anyone know if there will be Skirmish AI in the Alpha release to play against the computer?
I'm pretty sure there will be. Alpha-testers would "teach" AI if i can say so. And beta-testers would face a competitive computer enemy.
Doubt it, I don't see much point in doing all the AI prep work and 'training' when the game isn't even feature complete, nevermind the fact that it can't be properly/fully 'balanced' until it is or nearly is feature complete. Perhaps once the game gets closer to feature complete(and thus closer to Beta) we might see some AI stuff popping up, but until then the Alpha is more so about making sure the game runs on different systems and that features actually work outside of the Dev Environment. Mike
I think there *may* be a first form of AI. We won't train it, I guess we can post on a thread here in the forums what can be improved so that Sorian knows what the current problems are. Remember Supcom2 AI? It wasn't that bad, but at start, it spammed transports all over the place. Thanks to the community that was fixed. A very first form of AI, perhaps. But we won't teach it, nor will it learn from battling against us.
At best, we're likely to see an AI that needs testing, not training. IE, please report how & why the AI broke during your game. Not even balance / difficulty stuff, but simple things like the commander stopped moving, the AI never built X or Y, or the debugger reported the AI thread was raising exceptions all over.
Actually, both of you are wrong. There won't be any AI in alpha, at least not in the first version. They just aren't there yet as Sorian has quite a lot of other problems to solve first. And they AI DOES require training, but that training will not be performed in games vs real players, but in simulations of hundreds to thousands of situations under controlled conditions in an automated, not real-time process against previous versions of the AI. Thats at least true for the unit / squad AI, but chances are, that the strategic AI will partly use similar techniques. We might see unit / squad AI in the course of the alpha, but real skirmish AI, especially the strategic component, might take much longer. It could happen, that there won't be any competitive AI until beta or not even until release. The alpha will most likely only contain a very basic unit AI which can do little more than just pathfinding and aiming, but don't expect units to behave smart in any way.
Does no RTS AI get any training from its playerbase? I mean in any game, not just PA. That'd be a pity.
No, it doesn't. What you call "AI" is not really an AI, it's just a complex rating system based upon many rules. You can't really "teach" that thing new tricks, all you can do, is to slightly correct the parameters in the existing set of rules to improve the accuracy. And even THAT requires neural networks! A "basic" skirmish AI would most likely not even learn anything at all, but would be purely based upon static rule sets and clever heuristics. Actual learning would require genetic algorithms, but as far as I know, there is not a single working implementation of an AI based upon genetic algorithms, at least not outside of an laboratory and surely not in any game.
what i meant is, are there no active perceptrons within the neural network for an RTS AI? Are their weight values never changed after release outside of patches? I'm not that much of a fool. I know they can't learn new moves, but they can see that some of their own moves aren't winning, and use those less accordingly. also, this would count as a genetic algorithm learning AI, and it's been around for years. Not really a game per say outside of its highscore stuff.
Your example shows a genetic algorithm, yes. But not in the context of an AI. Last year I've heard a presentation about the use of genetic algorithms to create basic motoric algorithms for very simple, simulated bodies with as little as 5-10 joints, but that problem was still much simpler than creating an AI for an RTS. Right now, even the smartest, real AIs are just as intelligent as a fruit fly. Research is still years away from creating AIs which are capable of abstraction and without abstraction, an AI will never be able to solve complex problems, not even approximative. Feedback is usually disabled, there would be no point in running the expensive evaluation algorithms in a productive environment, after all you are using the neural network because it provides a both efficient and effective solution to the problem. If the network fails, then it is most likely due to a flawed input or broken implementation. But you can't fix that by adjusting the weights as you just mess up the whole network, this kind of issue needs to be fixed by adjusting the input filters and the implementation of the available moves which are connected to the output.
Bleagh, why doesn't MIRI hurry the f**k up and finish doing the math on real AI. They're doing such good work, pity it's not really understandable to me or anyone outside their field. Have you see their stuff? Eliezer Yudkowsky works with them, they're trying to make a learning AI that is also objectively friendly.
