When do you PWN the AIs in an SP game?

[Chaos Theory]

People can play SMAX well, can't they? Just how smart do you think people are? We don't analyze all of the myriad possibilities when playing, we estimate. For example, when expanding, we look at factors of production lumped together, rather than individually analyzing the nutrients, minerals, and energy. Combat is not a known quantity, therefore it's best to bring more than enough to bear. Terrain varies, but forests fix that.

You could argue that it has taken many players together a few years to come up with all these good strategies, but computers can cooperate even better. Even if we don't know how to make a good AI for SMAX, it's not infeasible. Just because lookahead is doomed due to the complexity of SMAX doesn't mean heuristics and learned behavior won't work.

[Dansz]

Perhaps the Turing test should be replaced by the SMAC test
Anyway, I think any computers capable of such things as learning are probably being used for more important things than games.

[Chaos Theory]

Personal computers are capable of learning. The block is more a lack of good software than a lack of sufficient hardware, and software, once written, can be copied infinitely. Of course SMAX, as a game, does not deserve much effort, but people like to spend free time on frivolous, but fun, endeavors.

[Skanky Burns]

Thats true enough, but then the software isn't there.
I just hope that Firaxis make a SMAC-like sequal so we can see Soren's efforts at this.

[Santiago_Claus]

It's been my observation for almost all games that strategy guides generally contain only very basic "strategies" and that all the truly useful information worms its way out of the fan sites.

This is probably reflective of game AI's, as well.

We assume that all our game-winning strategies were simply Easter Eggs placed by the designers waiting for us to find them, that Brian Reynolds meant for us to realize we could relentlessly troll fungus like prospectors panning for gold, or that beelining to IA and cramming the continent with crawlers was part of the designers' "vision" for a well-played game.

Maybe it wasn't. Maybe we today are better players than the designers were AT THE TIME THEY WROTE THE AI. Maybe our tried-and-true tactics hadn't been invented yet -- even by the game's inventors! It would be a tall order indeed to expect an AI which could exceed the skill level of its designers.

I'm quite sure that the designer of a game, once exposed to a certain strategy, could derive insights far beyond any of us, and could both play and program an AI accordingly. By that time, though, the game's already out.

[Chaos Theory]

And this leads straight to learning algorithms, particularly evolutionary computation. You basically write an AI that can be tuned via a variety of parameters and have a bunch of versions of the AI play thousands of games against each other. During these games, they could slowly discover such tricks as using crawlers, or taking advantage of fungus, whether or not the game developers had realized the possibility. It is entirely reasonable to expect an AI for a game that plays better than its designers can (though they should learn from it once it exists).

[Sikander]

Originally posted by Chaos Theory
And this leads straight to learning algorithms, particularly evolutionary computation. You basically write an AI that can be tuned via a variety of parameters and have a bunch of versions of the AI play thousands of games against each other. During these games, they could slowly discover such tricks as using crawlers, or taking advantage of fungus, whether or not the game developers had realized the possibility. It is entirely reasonable to expect an AI for a game that plays better than its designers can (though they should learn from it once it exists).

Btw I agree that brute force isn't the answer to this sort of complexity. How about instead of a bunch of computers playing one another to come up with new and interesting strategies, instead we have a bunch of players play against versions of the AI and those games are recorded and analysed in order to create better versions of the AI. Indeed a system like the one Deep Blue uses could be considered, where an AI is tailored for a particular player. Imagine a game that constantly updated its AI and made it available in a central repository. It would be immediately field tested by people playing the game, and as they figure it out it is also figuring out how to deal with their new strategies. New players could always opt for older / easier versions so that they can work their way up.

[Dansz]

Computer's playing against one another is a good idea IMHO. It's been done before with simple things like 'the prisoners dilemma' sort of stuff, but that was a long time ago so probably far more complex inter-computer gaming has been done since then.

I think almost certainly the strategies people have developed were not planned beforehand. That's why sports rules need constant reviewing, because people come up with new ideas that stop the rules working how they were intended. Either the SMAC rules need changing to stop people exploiting them, or far better, computers need to be programmed to use the same exploits.

[Santiago_Claus]

I hadn't considered neural net based AI as opposed to expert system AI. You're right, of course: I could write a chess program that played way better than I do.

The problem with a neural AI for SMAC-X are there are too many indirectly-related degrees of freedom in the game for the state of the art as it stands now to make much headway against the problem.

Put another way, it would be extremely difficult to "teach" an AI (or for an AI to "learn") to plan moves of units that don't yet exist then build those units to meet that design, whilst also taking into account the opportunity cost of building military over infrastructure.

But all-or-nothing presentations belie a middle ground; if there were a "combat AI" which made chess-like decisions and a "build AI" which ran off of scripts, the "combat AI" could maybe have a "pushback channel" in which it could demand particular unit types and an override priority against other builds, sort of like a "DefCon" level.

Would sectioning the AI's decision domains like that cut the degrees of freedom of each to a manageable size?

[Jamski]

But all-or-nothing presentations belie a middle ground; if there were a "combat AI" which made chess-like decisions and a "build AI" which ran off of scripts, the "combat AI" could maybe have a "pushback channel" in which it could demand particular unit types and an override priority against other builds, sort of like a "DefCon" level.

I think this is how GalCiv is working.

-Jam

[johndmuller]

The game already has a number of elements more or less quantified. For example, it has the four tech types and the various social engineering paradigms. Each faction is provided with predilections for particular combinations of these elements and they demonstrably effect the way the factions play the game. The weapons/armor/units have a field called 'mode' or 'plan', which seems to be intended to be used by the AI in some fashion. Presumably there are a number of general goals involved in the AI's decisions of what to do in combat, strategy and other stuff like terraforming (some of which obviously could have used more work). Anyway, my point is that there is a hell of a lot of the fundamental stuff already in there and it works pretty d@mn well, especially considering how non-specific it is.

As has already been pointed out, some of the most prominent failings, for example the AI's lack of enlightened usage of crawlers, were no doubt due to the designers being unaware of the possibilities - surely they could have programmed in some decent AI usage of crawlers, or at least made human usage less rewarding, perhaps by making them quick to defect to other factions, or to malfunction frequently enough to be problematic, say by delivering poisonous food or radioactive minerals to your bases. Other problems, such as the feeble terraforming, may have been due to last minute tweaking of things like the yield of forests, mines, farms etc, without sufficient playtesting and/or follow through into the AI terraforming routines. And there are also the obvious bugs, like the Stockpile Energy bug, which were just plain oversights.

One thing that is definitely lacking is any kind of results oriented evaluation and future behavior modifying aspects that carry forward from one game to the next. On a simple level, Tech choices and Unit selections could be rewarded/penalized for future games according to won/lost records; other aspects of the game might be more difficult to tweak, but even meat-ax approaches can zone in on useful strategies this way. The main key is to come up with the relevant stuff to reinforce or discourage, and insightful means of evaluating positions and results, all much easier said than done.

Then there is the look-ahead method.....

[Barinthus]

In other words, Firaxis didn't take their time to make sure the game was beta-tested properly and basically rushed it out the door as quickly as they could?

[Dansz]

Barinthus, doesn't that apply to all software, maybe even to all products of any kind? And then of course they follow microsofts example and sell add-ons that (arguably) should have been there in the first place.
I bet Firaxis isn't as unpopular though.

[Santiago_Claus]

Compared to other game companies I think Firaxis did an excellent job with their release version of Alpha Centauri. The infamous bugs everyone talks about are subtle enough that their slipping through beta was understandable. I wish more companies would release software as vetted as Version 1.0 of Alpha Centauri.

Firaxis' atrocious after-market support, however, was of the "hey, we're done, now leave us alone" variety; several well-known bugs went unfixed and in some cases even unacknowledged and remained so even through the expansion -- indeed, even unto today.

Nevertheless AC as released was about as bug-free and complete as any game in recent memory, with the possible exception of the impressively polished Warcraft III.

(Blizzard then turned around and obsessively and repeatedly patched, and those patches had a tendency to break things requiring further patches -- one of the patches made the single-player campaign almost impossible, requiring another patch to un-patch that part of the previous patch! But Blizzard's slap-happy patching is an off-topic for another time ...)