Heading - New Scientist Technology Blog

A technology blog from  Heading - NewScientist Blogs

Friday, October 05, 2007

Neurotic software has a winning personality

They say the meek shall inherit the earth, but these experiments with emotional computer programs (pdf) suggest it may actually be the neurotic. And that they'll probably take it rapidly by military force.

The Austrian researchers want games to be more engaging by having emotional, not just coolly calculating, computer players. Instead of just challenging your rational planning and decision skills, you'll have use your emotional intelligence too.

They created aggressive, defensive, normal and neurotic versions of the AI software in the war strategy game Age of Mythology, drawing on "the big five" emotional dimensions to personality recognised by psychologists.

The bots are able to switch between states of pleasure, pain, clarity, and confusion in response to events. The strength of particular emotional changes is related to the overall personality.

The neurotic bot was more likely than the others to distort hard facts about resources - like the amount of timber around - and flip between extremes of behaviour. And it was better than the rest.

Each bot took on the game's default AI seven times. Both the aggressive and neurotic bots won all their matches - but the neurotic did it faster. On average by around 25%.

There's no indication of whether being top bot boosted the neurotic program's low self-esteem. The researchers say their result suggests a neurotic strategy gives an advantage, and are going on to pit the emotional bots against humans. It'll be interesting to see if neurosis still wins - and if humans can tell the different personalities apart.

Tom Simonite, online technology reporter

Labels: , , ,

Comments:
All comments should respect the New Scientist House Rules. If you think a particular comment breaks these rules then please let us know, quoting the comment in question.
I suspect the neurotic AI my prosper by virtue of its unpredictability.

Other AI's make decision largely based on evidence which in turn means that another player can come to a similar conclusion based on the same conditions and predict what the AI will consider the optimum strategy.

The neurotic AI,however, attaches arbitrary fictional weights to various inputs making it difficult for others to predict it actions based on the same inputs.

You see the same thing in chess in which a novice player defeats more skilled one by virtue of making sub-optimal moves the skilled player does not even consider.
Hah! You see? You're all out to get me, you all think I'm weird, I can tell... but I'm the winner. Here's the proof.

Or am I? Is this just a trick? Oh god, is anyone actually reading this anyway?

Aaaaahhhh
This article implies that facing the built-in AI in a RTS game is anything like facing another human.

This really isn't true. So, what this article should really be titled is "Neurotic software is better than Age of Mythology's AI", because that's all it proves.
I have played many chess tournaments and don't recall ever seeing a novice defeat an advanced player by making suboptimal moves. It just doesn't happen. Big upsets are rare and invariably the result of impatience resulting in a serious mistake on the part of the advanced player.

However, poker is another story altogether. The unpredictablilty and downright craziness of a neurotic player such as Stu Unger can make mincemeat of more technically adept players. Disciplined, calculating players often find themselves on tilt trying to predict holdings and moves. Brilliant and neurotic is the sure path to success in such games.
The screen shot you have above is not from Age of Mythology but, age of Age of Empires 3.
Of course it proves their AI is better than the Mythology's AI, its called a baseline genius, and playing a human (read: incinsistant) player doesnt' give them that, that would be the NEXT phase of testing
The default AI in AoM is under-aggressive (compared to a normal player) and does not attack or defend in sufficient organized force.

In other words, it stinks.

Pit the neurotic AI against all the other types they made and see how it does.
It demonstrates the old advertising agency adage.

The craziest guy in the room, runs the room.
This test was really just who beats the 'default' AI best. I would really want to see how the neurotic fares against all the other AI's. It could be this is just a 'rock, paper, scissors' scenario, where neurotic beats default, but one (or more) of the others beats neurotic.

I'm also curious how it does versus human opponents. It would be difficult to quantize those tests, due to the large number of uncontrollable variables with humans, though.
The skeptic in me thinks that they've actually demonstrated something different than they think that they did.

It's well known that AI for RTS is notoriously weak and that the problem of creating an expert system for playing exploration/expansion games is not as well understood as say creating an expert system for playing chess.

I hypothecize that the default strategies which thier AI uses are far from optimal, and that the Neurotic AI is successful not because it is Neurotic but because it most frequently departs from the suboptimal strategy which it is programmed to employ.

In otherwords, all they've really demonstrated is that they don't know yet how to teach a computer how to play.
"I have played many chess tournaments and don't recall ever seeing a novice defeat an advanced player by making suboptimal moves. It just doesn't happen. Big upsets are rare and invariably the result of impatience resulting in a serious mistake on the part of the advanced player."

Shannon merely said "more skilled" and not "advanced". Certainly less skilled players and teams upset more skilled players and teams all the time. But everyone knows that a little league baseball team has no chance against a pro team just because they pitch a few fits instead of fast balls. But where things are close it may be valuable to open with a rook pawn from time to time.

So I agree with Shannon and think that Spassky-Fischer is a good example of the advantage of being unpredictable and at least looking neurotic.
This so explains Microsoft's success.
Another thought is that the neurotic AI may have merely built up more quickly at the expense of longevity. One wonders how it would have faired in a situation where it had to actually manage its resources...

That all said, this may be yet more evidence that a good defence is a good offence...
i'm wondering here..
isnt it rather the self reorganization potential

A neural network, which is trigered earlier in a neurotic net. then other types of nets, in this perspective i would rather call it the creative net, while the offender net is probaply the most logic one. Like people some are straight thinkers others are creative.
I think it's rather clear in the content of the article that the results only apply to the default AI. I think it will be extremely interesting to see where the research group takes this software. As others have mentioned, seeing the results of the various emotional programs pitted against each other, as well as matches against human opponents, will be fascinating.
This reminds me of the 'tit-for-tat' strategy of AIs in a previous AI v. AI tournament based on solutions to the Prisoner's Dilemma. Of the many strategies tried 'tit-for-tat', a simple strategy of responding in kind was the most successful. see here

I would guess that the neurotic personality of sudden mood swings and responding violently to an attack would be like tit-for-tat.
As someone who has written AI for a retail strategy game I'll tell you this, the best AI I ever wrote was what you would call totally neurotic. Not totally random though. It would "go for it" even if there was a high level of risk involved in the move. This tended to take players by surprise and they would find themselves on the defensive. This seemed to work the first few times they would play any scenario, but then they would get the hang of what the AI was doing and then would tend to crush the AI. I expect this is typical of any game AI though.

One guy even complained that the AI was horribly bad, but he would run the game with full intel to study what it was doing before he tried to actually play. pbhttt.

Doug
As a Civilization fan since the days of version 2, your report on the results of the study is personally intriguing. Before Civ IV, the most aggressive players, human or AI, dominated the game. As it has evolved though, strategies have become much more complicated, especially for those who play "fair" and don't exploit tech flaws or cheats. In my experience, games with static conditions like chess will be dominated by prudent players, but the more wide-open the conditions of play, the more an irrational strategy - or lack thereof - tends to pay off with victory - or at least the cold comfort of mutual destruction. I'd love to test if DARPA has run advanced game simulations to test if the neocon "might makes right" strategy is best in a post-Machiavelli world. My hypothesis is that the Straussians using the world as their proving ground might be better advised that the world has become a helluva lot more complicated since the Dark Ages. Peaceful but strong and richly dimensional cultures like India and Egypt dominate. My best run as India produced a diplomatic victory without ever going to war, and that is REALLY hard to do in any version of Civ. As far as a real-world application, irrationality quickly leads to isolation and disaster, as N. Korea illustrates. Just doesn't work in the modern world like it did before the age of mass communication. Crazy like a fox is another story....
All the personality AIs were significantly better than the games AI and won at least 6 of 7 games in different scenarios. "Neurotic" was simply a label for tendency to switch between extreme playing styles, i.e. aggressive and defensive. This is simply a prudent strategy most successful human players employ in these strategic games. For example, you initially focus entirely on defense, then you switch gears and focus entirely on offense (is this really "neurotic"?). This will usually result in a fairly quick kill. The limits of the alternative personalities are self evident: A 'normal' personality will result in a game that drags on, as decisive battles needed to end these games will be avoided. The same goes for a 'defensive' personality. Likewise, if one is constantly very aggressive the game will drag on as well since you will be sending your troops there in a constant trickle rather than waiting for a decisive batter. Thew only way an aggressive personality will win is if there is a huge discrepancy in skill, which one to completely overwhelm their opponent right off the bat. But clearly there was not such a massive difference in AI in this case.
I'd like to see the neurotic software pitted against some of the better players in the RTS market, or even other AI from different games. Especially games with a larger fanbase, such as Starcraft.