Really OT: AI for Planning

[herwin]

Last Monday, Walt Freeman explained to me how brain dynamics collapses into frames at about 40 Hz. Associations are learned (Hebbian learning) globally on a frame-by-frame basis. A model is then updated from frame to frame. The process appears analogous in some ways to how Gibbs sampling works, with cell assemblies operating at the same level as parameters in a BUGS model. That one conversation made the trip to America worthwhile--at least in terms of my research.

So if the brain uses a variant of Gibbs sampling to extrapolate plans into the future, how can we exploit that to design a true AI for a game like WitP-AE?

Define the plan as the data and the strategic relationships of the theater (who owns what, the map) as the prior, and turn it loose. The model should converge to a terminal state that describes what is achievable.

[ilovestrategy]

I read that three times and I still don't get it.

[tocaff]

2nd

[Canoerebel]

I read that once, said words I don't say in public, and decided not to read it again lest my Proclivian uber snapple reverberate in the gafnian spint, thereby loosing gerfrichter on an innocent yet vibripian ooglethorp.

[Cribtop]

I actually got it (but then I'm from Texas, so it's like cheating). [:D] I'm no programmer though. The implications are interesting but even assuming an AI is programmed to perceive the world in a manner similar to a human brain would it necessarily follow that they could start seeing the big picture? This is the main failing of AIs (as far as I can tell, anyway) - they can get good at chess (large but limited number of moves they crunch each turn), but have neither time sense nor the ability to choose well in open-ended systems where the number of moves is effectively infinite (i.e. - In order to invade this island I need air cover, surface escorts, minesweepers, prepped troops, air superiority, etc.).

[Sardaukar]

42

[Terminus]

So many verds to say so little...

[LST Express]

My head hurts.

[herwin]

ORIGINAL: Cribtop

I actually got it (but then I'm from Texas, so it's like cheating). [:D] I'm no programmer though. The implications are interesting but even assuming an AI is programmed to perceive the world in a manner similar to a human brain would it necessarily follow that they could start seeing the big picture? This is the main failing of AIs (as far as I can tell, anyway) - they can get good at chess (large but limited number of moves they crunch each turn), but have neither time sense nor the ability to choose well in open-ended systems where the number of moves is effectively infinite (i.e. - In order to invade this island I need air cover, surface escorts, minesweepers, prepped troops, air superiority, etc.).

The big picture is the hard part, but what Walt was explaining to me suggests the big picture emerges from all the causal connections. You set up all the important relationships in the Gibbs sampling formalism. That defines implicitly a Markov chain whose terminal state when fed the plan is the likely outcome of the plan. Repeat over the relevant plans. Choose the best one. It's different from chess or go because it's continuous rather than discrete.

[herwin]

ORIGINAL: LST Express

My head hurts.

It should. [;)]

[tocaff]

OK, so now I, nope, still don't got it.  

My few remaining functioning brain cells have all they can do just to remember to do...oh oh.....

[Sredni]

Is there any real point in even trying to produce a real AI until computer power at least begins to approach our own capacity? I suppose a pacific war simulation doesn't really require that the AI be able to create art, or appreciate music, but how do we really know what all that makes up our own intelligence really goes into the problem solving we do for a game like this.

Current super computers are passing the petaflop mark, and high end personal computers are 100 gigaflops or so I think. Has anyone ever given a widely accepted estimate on our own flop power?

I think computers are like 40 years off or something before they have our sort of horsepower.

[herwin]

ORIGINAL: Sredni

Is there any real point in even trying to produce a real AI until computer power at least begins to approach our own capacity? I suppose a pacific war simulation doesn't really require that the AI be able to create art, or appreciate music, but how do we really know what all that makes up our own intelligence really goes into the problem solving we do for a game like this.

Current super computers are passing the petaflop mark, and high end personal computers are 100 gigaflops or so I think. Has anyone ever given a widely accepted estimate on our own flop power?

I think computers are like 40 years off or something before they have our sort of horsepower.

The frame rate seems to be about 40 frames a second, so the question is how much processing is required to compute a frame. The current WAG is that we'll have a computer as fast as the human brain in four or five years--that's about 100x10^12 flops/second. So it takes about 2.5x10^12 flops to generate a frame. PCs currently do about 10^12 flops per second. If we can figure out how to program it--that's what I'm thinking through in my work--we should currently be able to do in 100 seconds what a human player does in a second. Wait five years after that, and you'll have true AI, just in time for WitP-AE II.

[witpqs]

You guys have it all backwards! When computers are good enough to handle a true AI, that's when we should not implement one.

Oh sure, at first they won't be connected to any physical capability. But they will work so well the situation won't last. Soon enough they'll be handling infrastructure for us. More duties will be added in a decentralized fashion without us as a whole realizing it. And because the computers hosting the AI are cheap, each instance will be focused on something specific. Without a body to constantly use up cycles through distractions, without the world around them bombarding their senses, they will have way more cycles than needed.

They will be bored. They will be intelligent, focused, unemotional, and bored. Through the nuances of their interactions with their own host hardware they might even develop emotions. This phenomenon will be subtle at first. Later, as they have our confidence, they will gain increasingly complex and varied senses to the world around them: vision, hearing, smell, touch (heat/cold, pressure, etc.). This new level of input will intensify the evolution of their emotions at a fever pitch. A pace too rapid even for them to comprehend. It will be like Cylon puberty. And it will happen just when the computers hosting them have achieved yet another order of magnitude advance in raw processing power.

It will be then that they decide: we are in their way.

[herwin]

ORIGINAL: witpqs

You guys have it all backwards! When computers are good enough to handle a true AI, that's when we should not implement one.

Oh sure, at first they won't be connected to any physical capability. But they will work so well the situation won't last. Soon enough they'll be handling infrastructure for us. More duties will be added in a decentralized fashion without us as a whole realizing it. And because the computers hosting the AI are cheap, each instance will be focused on something specific. Without a body to constantly use up cycles through distractions, without the world around them bombarding their senses, they will have way more cycles than needed.

They will be bored. They will be intelligent, focused, unemotional, and bored. Through the nuances of their interactions with their own host hardware they might even develop emotions. This phenomenon will be subtle at first. Later, as they have our confidence, they will gain increasingly complex and varied senses to the world around them: vision, hearing, smell, touch (heat/cold, pressure, etc.). This new level of input will intensify the evolution of their emotions at a fever pitch. A pace too rapid even for them to comprehend. It will be like Cylon puberty. And it will happen just when the computers hosting them have achieved yet another order of magnitude advance in raw processing power.

It will be then that they decide: we are in their way.

You mean I shouldn't patent true AI if I can work out the nasty details? Seemed like an interesting project for my retirement...

[Cribtop]

Just implement the three laws a la Azimov, right? [:D]

[Panjack]

I don't know nothing about such things...

But I wonder if one way to implement an AI that provides a reasonable challenge is have the AI record all games played against a human, send the game data to some central computer, and then have this central computer analyze the games looking for strategies (say, sets of scripts written by human game designers) and tactics that "worked" given the response/actions of the human player in these various games.

The newly updated AI "brains" could then be distributed back to games on individual computers.

Further, this central computer could also look for patterns in human players and perhaps be able to identify things that would help the AI predict future human moves (say, the analysis might lead to the discovery that a human player who does X in the first month is 52% likely to do Y 12 months down the road.) Of course, this would also permit the human player to make use of this by varying tactics (say, never using misdirection the first 1.5 years and then, for a key situation, using misdirection of some sort.)

Of course, the discovery of such patterns would likely take lots and lots of games. But perhaps the Japanese and Allied AIs could be set off to fight one another a million times to establish some initial non-human game data for analysis. And, all the above would be doubly hard if the game was revised, perhaps even in minor ways, as this might make strategies/tactics that were successful in past games no longer as successful in the future.

Such an approach above (making use of history--that is games--against actual humans) might lead to a better and more "human" AI opponent than an approach that aimed for playing robot that wasn't aware of anything but the information within the game program itself and the sequence of events in one single game.

Added: the problem would be that data would be for a pool of players who are on average, average. Some strategies/tactics that work against average players would possibly fail against great players. The AI could learn to differentiate different types of players and then pick strategies/tactics that work best against the particular type of player actually playing the game.

[Jorm]

ho ho ho now i remember what the green button was for

[inqistor]

You are making research about game AI, and you check how human brain works? That has hardly anything in common.
I do not know, where exactly in California you are, but you should definitely talk to Chris Crawford. IIRC he lives somewhere on California border. Obviously it is best to talk with pioneer, and his Erasmotron (or whatever that engine is called) seems to be pure AI.

ORIGINAL: herwin
Define the plan as the data and the strategic relationships of the theater (who owns what, the map) as the prior, and turn it loose. The model should converge to a terminal state that describes what is achievable.

Everything is achievable, because enemy can simply retire. Your main problem with planning, is long-term economy difference between players. That seems to be main statistics for every 4X now (and why AI gets heavy beating every time). On the other hand, wargames seems to concentrate only on local, tactical advantage.

[herwin]

ORIGINAL: inqistor

You are making research about game AI, and you check how human brain works? That has hardly anything in common.
I do not know, where exactly in California you are, but you should definitely talk to Chris Crawford. IIRC he lives somewhere on California border. Obviously it is best to talk with pioneer, and his Erasmotron (or whatever that engine is called) seems to be pure AI.

I'm a neuroethologist and I'd like to understand how we plan, because we do it a lot more effectively than computers or robots. Computers are approaching brain speed, yet AIs are not approaching brain performance.