'Neurotic' is Best RTS strategy 186
An anonymous reader writes "Austrian researchers experimenting with adding emotion to game AI say that 'neurotic' software is best at RTS. They developed aggressive, defensive, neutral and neurotic bots to play Age of Mythology, based on psychological models of emotion. Neurotic bots beat the standard game AI every time and faster than the other personalities."
Welcome to Lords of the Realm 2! (Score:4, Interesting)
I discovered that a hardcore neurotic kind of strategy worked well in Lords of the Realm 2 when playing with my brother. He didn't care, and would rather have the game over quicker than not, so when we started the game he immediately spent all his resources on getting weapons and a huge army, and within four turns or so had come over and whooped my ass. Every single other aspect of his kingdom was in shambles, but he had the element of surprise, and that's all that ended up mattering.
I'm thinking the AI would think something similar to me... "Surely he won't try that. If he fails in his attack, he'll just fall over on his own accord in a few turns." Unless he doesn't.
"Best" (Score:1, Interesting)
Makes sense (Score:3, Interesting)
faster is better? (Score:3, Interesting)
Faster is better now? Then why did they bother to code the defensive personality?
But who cares about some real-time strategy game? (Score:5, Interesting)
To get a meaningful result they'd need to test the different programs against experienced, intelligent human opposition. Or better, stop messing around with real-time strategy games and design AI for a game whose rules are already well-known. If a 'neurotic' or 'emotional' player program starts beating the 'purely logical' computer engines in chess, then I'll take notice. We know that the existing AI for chess is quite good (and there is a choice of several strong engines to test against) so any advance over that is likely to be genuine and not just exploiting obvious flaws in some existing program.
Re:Result is specific to AoM? (Score:3, Interesting)
Having watched some amazing starcraft players, neurotic sounds about right.
crazy leaders? (Score:3, Interesting)
Re:Result is specific to AoM? (Score:5, Interesting)
Reading the article, (which is freaky low on detail) it seems more like "Neurotic" in this case is meant to signify a lack of caution. Aggressive won every match, and neurotic won every match, but neurotic did it faster. This suggests that irrational risk taking (the article mentions that the AI skews its internal numbers about how many resources it thinks it has) can beat a more cautious opponent.
In both cases it seems clear that aggression carried the day, and that the only real difference was that the AI that lacked caution won faster. To me, that suggests a big problem with the regular AI, because that lack of caution is usually pretty easy to exploit...A counterattack on a resource gathering operation would leave the crazy AI crippled, due to low reserves. The regular AI's counterattack algorithms must be pretty weak, or it's build order is too cautious or something.
I'd love to see a better description of the AI programming.
Re:Result is specific to AoM? (Score:3, Interesting)
Years ago I remember reading about a study which "proved" depressives were more cognitively accurate, by setting up a task which was fixed so that the subjects always failed. The depressives of course recognized this much more quickly than the normals.
However, that said I think there is still an interesting point here. The neurotic profile may exaggerate the situation, but at least it reacts to it, as opposed to inbuilt tendencies toward being aggressive or defensive all the time. One possible benefit to emotions is to engage survival behavior early, before perfect information on a situation is available.
On the other hand, the way to defeat a neurotic is to deceive him into misreading a situation. Once that happens, he will not adjust to contradictory information.
Re:Makes sense (Score:4, Interesting)
Believe it or not, the old AI's in Age of Empires, with no sense of retreat, were harder to fight as they'd send their forces at you non-stop. The game was almost completely about whether you can build an army faster than the AI because the AI would not hesitate to send his entire army after you as soon as he developed it.
Bad measures of AI (Score:4, Interesting)
The default AI in most games is terrible - even just writing a "do-random-stuff" AI would probably beat the in-game AI 20-50% of the time (provided you put in simple anti-suicide routines, like not using up all it's available funds etc.). Most AI in games relies on the fact that it knows everything that's going on (including exactly how long until their next unit is built, how many pixels you are away before it can fire on you, how much gold it will have by then etc.) and will generalise EVERYTHING (i.e. it'll be in "attack" or "defense", "hard" or "soft", "co-operative" or "go-it-alone"). Most games have a variety of "sliders" on the AI and the games-makers tweak them either randomly, in steps for each more difficult level or according to a pre-built AI "profile" (e.g. cautious but fast etc.).
In some games, that's more than enough to give anyone a challenge, at least until they are nearing the end of the game's useful lifetime. Snooker/pool games spring to mind. You won't beat a "top-level" AI on a snooker/pool game. It knows exactly where everything will go, even several "moves" in advance if necessary and can play a perfect game if required.
RTS's though, are much harder to simulate. Yes, there are a lot of factors involved in the creation, strength, durability, mobility etc. of units but at the end of the day it's a military tactics game. Pixel-perfect positioning of a nice ambush will keep the computer in an endless loop of "attack, run away, heal, attack, run away, heal".
I've not played AoM much, I'm an AoE2 fan personally, but the AI was amazingly easy to overwhelm with just a simple early-game rush, confuse with an impenetrable fortress hiding some long-range weapons and particularly predictable when it comes to individual AI tactics.
All AI's are predictable to a point in mass-market games - you can always "learn" to beat the AI in any particular game. Granted, it may be hard to do, it may be different to other similar games, but there's always some point at which you "know" what it's going to do.
It seems to me that, given that, an AI that is very "jittery" and over-compensates might beat the in-game AI in some games. However, on others, even in the same genre, it would get trounced. The "researchers" are assuming that the in-game AI is somehow a good approximation of a "neutral" player. They are also assuming that they have programmed each type of AI without any glaring logic holes in their tactics and that they are all equally matched in terms of capabilities. A cautious AI would win over a boisterous AI in only 50% of games.
More importantly, it's only a test of AI programming skill, not what "personalities" are trying to be reflected by the coders.
Re:Result is specific to AoM? (Score:4, Interesting)
I'd think that in a 3 person battle a neurotic AI would be at a great disadvantage because the style of "to hell with the consequenses, charge!" game play might win against 1 but not with a 3rd party. They'd jump in take advantage of the neurotic side when they had no reserves left and had spent themselves fighting the 2nd opponent.
addiction (Score:4, Interesting)
Well, sort of seriously, anyway.
The cultural tradition of women getting their way stems, in my opinion, from the cultural reinforcement of addictive tendencies in men. More specifically, addiction to sexuality. While the male sex drive is strong, cultural influences encourage even more slavery to this impulse, and further incline one to view a low sex drive (or even just a stoic level of self-control) as a lack of masculinity, or simply put, as weakness.
The end result is that men adopt a strongly sex-driven persona which in turn gives their women great control over thier behavior.
In other words, our notion of horny=manly sets us all up to become p-whiped.
The door swings both ways. Biology + cultural reinforcement inclines women (at least American women) to want romance (especially to be seen in public with a man who is showering affection on her). Learn to grant and withold that, and you can start getting your way too.
Re:But who cares about some real-time strategy gam (Score:4, Interesting)
But thats just it. Chess allows only for the "Next Best Move". Playing an illogical move only results in the player playing it to loose because it puts them at a disadvantage and the logical computer simply knows the counter moves anyways for your worst move.
As in...
A logical AI assumes you'll play the next best possible move, but if you play the next possible worst move you are in a worse position and the AI simply knows the next best move and plays for that, but if you still keep playing the worst possible move you will only end up loosing faster.
In that regards, a logical chess program would be an AI or human who plays non-logically.
However, the reason why an RTS is important is because Chess is a limited game to a certain subset of rules that a computer can brute force all possible best moves.
However, in real world combat situations, there are no set definitions of strategies because you are simply allowed almost infinite possibilities of winning.
Lets say we take a human pilot or an AI pilot in actual Fighter combat in the skies (we'll see this scenario in the next 20 years) and pit them against each other in a real world situation. A logical AI would understand what the next best move is and the pilot will have an idea of what a logical AI would do.
However, the human pilot might do something crazy it knows it can throw off the AIs strategy like flowing into a nearby storm cloud or perhaps into a dangerous maneuver through a canyon or city landscape (under bridges and between buildings) which might throw the logical AI off.
After a while, a human pilot would have a general strategy with dealing with an AI that didn't adapt. He would know how an AI would react and be able to defeat it without too much effort.
Now a completely crazy AI would basically confuse the human and also other AIs who assuming the other AI was going to do in its next best move. Since in the real world (and in RTS) there are almost infinite combinations of what you can do in real combat, being unpredictable really helps win battles.
But like I said... Chess only has a limited set of moves. I would be an illogical AI would do far better at a game of Go than his logical counterpart.
is "psychology" what matters here? (Score:2, Interesting)
That said, from the introduction of the presentation one can see that the real goal of this effort is to create bots that *people enjoy playing against*. That's probably only loosely related to the absolute strength of the opponent, and it makes complete sense that it would be thrilling to be up against an AI that can suddenly just "take a flyer" and surprise you.
Re:Bad measures of AI (Score:4, Interesting)
However, I can say with great certainty it is not only possible to enact emotion in a cognitive system, but is being done right now. I'm doing it and developing real software systems employing it. The standard computer model of emotion in computing, called the OCC Model, is partly wrong. It misses what's really happening in humans. I've developed a more correct model that works very very well and probably matches the mechanism people use. I haven't published it. Why? Because some of my key competitors are Google and Microsoft. (Yes, Google's working on AI, shades of Skynet, eh?) Anyway, it is far easier to build systems that accurately have and express emotion that ones that can read human emotions. In other words, having and expressing (output) are easy enough, reading deep emotion in others (input) is much more difficult.
A few ending remarks. A lot of people are working on not much more than toy AI, and I've read some DoD-sponsored papers that are so far off base they are sad. I believe the correct approach combines both symbolic and analog AI (NNs) in a new way, and that we can create reasonable emulations, if not parallels, of human cognition. But they must come from a decent merging of psychology, sociology, and computing science. I've been working on the right path, a very productive one, charting a new course, and am writing what is currently a 5 volume book set I'd like to become the 'Knuth' of Synthetic Intelligence development. It should change the face of gaming and a few other things. Finally, I'm currently trying to emulate neuron-based systems in Erlang, by the way. Boy, is it parallel. I think that holds a lot of promise.
Re:Welcome to Lords of the Realm 2! (Score:3, Interesting)
The difference is that games have an end, whilst short of global catastrophe, real life does not. If a group is successful to the point of eliminating all other groups, it will soon turn on itself because the values that brought it to that stage demand it. If, for example any particular human group you choose such as Christians, blacks, whites, whatever, successfully elimated all other groups on the planet, you'd soon see it split into new, struggling factions. Because objectively, there aren't any differences between these groups significant to prevent working together. The differences are superficial or self-imposed. So there's no reason to suppose that new differences wont be found to justify another round of global supremacy. Co-operation and peace breaks the cycle. Winning never does.
Re:Result is specific to AoM? (Score:3, Interesting)
A couple of bunkers could wipe out the sort of zealot rush that is almost impossible to stop without bunkers chock full of firebats. At it's heart, its an AI problem. The AI chooses a target it's unable to attack, and then is unable to drop it. IMHO it's one of the oldest, and worst, bugs in strategy game AI. It's easy to exploit.