Could Deep Blue play a smarter AI ?
- Thread starter NapoléonPremier
- Start date
- Status
Zombie69
Emperor
- Joined
- Nov 22, 2005
- Messages
- 1,902
warpus said:
Easy, pit different AIs against one another, and then from the pool of different AIs you have, use a genetic algorithm so that AIs that won more games get to "mate" more and have a bigger percentage of their "genes" (probably the weights of the nodes in the neural net) represented in the next generation. Do this over a few hundreds of generations and you can get something really good.
That's how i did it anyway when faced with a similar task in an AI course that i took. It works very well too.
Dusty Monkey said:
Dusty,
Please, please, this just isn't important enough to get upset about!
I guess I'm not going to 'make up my mind' about computers typing in opening books in computers vs. humans. This was a HUGE debate in the late 70's and early 80's when people first started talking about it -- given that people come out on both sides I view it as an open debate.
When I play against the chessmaster program, it doesn't matter. The fact that the chessmaster can play openings well makes it a better educational tool. It does analysis and helps. However, I'm just a club player so playing a computer is just recreational.
When it comes down to 'Human vs. Machine', especially inthe 70's, most of the programmers felt that they could write the perfect algorithm. Many of them felt their computers could win through pure calculation. As with Levy, they were surprised at how good a REAL LIFE grandmaster was.
So, many of the PROGRAMMERS THEMSELVES thought it was cheating to type in opening books. They wanted to use their algorithms directly to defeat players. Many of them were shocked at how terrible computers were at openings. Of course, their computers fell behind the others.
Now, in postal chess, a human can use opening books. So, is the correct analysis human vs. computer over the board, or is it human vs. computer in postal chess? Many chess players are not convinced one of them is right.
The issue of whehter computers using opeining books is cheating is still debated at chess clubs I attend.
I'll take your word for it that others are better than Kasparov for fighting computers. I talked to A LOT of chess programmers, who felt their only competitiion was Kasparov. They may have been referring to total rating, not direct play -- I wasn't really asking them that. (Of course, a lot of this was before Krammnik entered the scene. Anyway, I think he's been a disappointement since beating Kasparov. Where would we put Topalov now?)
Anyway, I think we are hi-jacking this thread. It REALLY isn't worth getting angry over. If I'm saying anything that is really bad, I apologize.
I'm not going to write about this any more.
Best wishes,
Breunor
friskymike
Listening to:
There are still some very stupid things the AI does that could be easily improved without too much effort. Play a game with the AI on your time and watch what they do, or read these SG games:
Rat12 - Team Challenge AWN
Rat 13 - 3 v 3 AWM
Rat12 - Team Challenge AWN
Rat 13 - 3 v 3 AWM
5cats
Warlord
warpus said:Why not? Humans are after all biological machines. What would prevent other machines (be it hardware or software) from being sentient as well? [/quote=warpus]
Oh! I so hoped someone would rise to the bait!
By simply copying the human brain, (wetware) we have not 'created' anything. Think of it this way: The Mona Lisa is an amazing painting, but a photocopy of it? Not so amazing. Even if we make wet computers, the software is just that, software.
And zombie69, there's "spooky actions at a distance" also called "non-locality" which has proven that something can transfer information at a speed faster than light. Einstein's best theory, IMHO. The "clock in the box" experiment has been done, and it defies the speed of light.
AutomatedTeller makes a good point, multiple victory conditions make do AI programming a bit harder. Unlike chess, or even Go...
And just because past scientists have been wrong, doesn't mean some things are not impossible. I'm of the opinion that true Artificial Sentience is a Holy Grail. It's nice to strive towards it, but it can never be achieved.
Dusty Monkey
Warlord
- Joined
- Feb 20, 2006
- Messages
- 110
Breunor said:
Whos upset????
Are you one of those guys who sees big letters or capital letters and thinks "oh dear!! this guy is yelling at me!!!"
Thats one misconception an A.I. wouldn't have
Pawel
Daimyo
- Joined
- Mar 18, 2003
- Messages
- 625
Dusty Monkey:
Don't you see that he is looking for an opportunity to write more about chess? If that's what an AI would do, I must say it would be a quite clever one!
friskymike:
The good thing is that when they release the SDK, there will be an opportunity to fix those simple things...
Don't you see that he is looking for an opportunity to write more about chess? If that's what an AI would do, I must say it would be a quite clever one!
friskymike:
The good thing is that when they release the SDK, there will be an opportunity to fix those simple things...
A few people have made the point that having multiple victory conditions, with each of which requiring a different strategy, makes it hard for the AI.
Why not amalgamate all the victory conditions into a points system? This would be rating the civ as a whole, rather than simply setting a few benchmarks, one of which has to be reached.
Thus you could win by being best at nothing, but brilliant at everything. It makes more sense, and makes Civ more of a strategy game. You'd actually have to make value judgements: this will help the war effort, but could I get more points by being more balanced?
At the moment you choose a victory condition and then you're not playing civ, you're playing 'conquering armies' or 'press return and let culture grow'. There's not much tactics; you've chosen your strategy, and that's it.
Having many games in one (each victory condition representing a different game) is fun, but I'd rather have strategic decisions such as whether I should abandon my wars and build culture, or conquer a bit more etc.
Why not amalgamate all the victory conditions into a points system? This would be rating the civ as a whole, rather than simply setting a few benchmarks, one of which has to be reached.
Thus you could win by being best at nothing, but brilliant at everything. It makes more sense, and makes Civ more of a strategy game. You'd actually have to make value judgements: this will help the war effort, but could I get more points by being more balanced?
At the moment you choose a victory condition and then you're not playing civ, you're playing 'conquering armies' or 'press return and let culture grow'. There's not much tactics; you've chosen your strategy, and that's it.
Having many games in one (each victory condition representing a different game) is fun, but I'd rather have strategic decisions such as whether I should abandon my wars and build culture, or conquer a bit more etc.
5cats said:
At the risk of getting very off topic. I was under the impression that the non-locality experiments didn't actually violate Einstein's theory on the grounds that although some information is being sent faster than light, the rules of quantum mechanics seem to suggest it can't be reconstructed at the other end, thus making it utterly useless (unless we have a non linear Schrodinger equation but I think that is not a discussion for a Civilization 4 forum
).That isn't to rule out the possibility that Einstein couldn't be proved wrong (certainly quantum mechanics and reletivity don't really fit together at the moment), but I think rumours of his demise have been greatly exaggerated...
The human brain is a machine. If you copy it, you create a new machine. The fact that it is not created from scratch or that it ain't "amazing" is beside the point.5cats said:
If I'm an engineer and have access to a very powerful combustion engine, I may be tempted to copy its design. I may not understands all its details, but I "created it" anyway. After that, I can modify it at will (by "tweaking" the copied design).
In the same way, we could copy the human brain and then "tweak" it to make it better (more intelligent, whatever that mean). I don't see why it "wouldn't count".
What's wonderful with the human brain when discussing AI is that it is a proof that such machine exist. I know it is possible to have an intelligent machine because I have an example of an intelligent machine. That's already much more than many other machines where we don't know if they are actually possible or not (e.g. an anti-gravity device). For AI, we know, because we have an example of a working prototype.
Edited to add: I'm using "machine" here in the broad sense of the term. Obviously, if you equate machine = implementation of a turing machine, i.e. a computer as we know it, maybe AI is impossible in this context. But I don't see why we should limit ourselves to this subset of all possible machines.
warpus
In pork I trust
Zombie69 said:
If you really wanted to use a neural net to implement this strategy, you'd have to repeat the process MILLIONS of times to get anything useful.. and even then, the neural net would have had to been designed properly.. and in the case of civ4, that'd be hard. What do the nodes represent? How do you weigh the connections between the nodes?
Don't get me wrong, this strategy would work wonders for something a lot simpler - I'm guessing that the problems you tackled in your class were just that - very simple problems (was it checkers or something similar?). Look at humanity - we're a very complex organism. It took over a billion years of evolution and natural selection to produce us in our present form.
A couple hundred iterations just won't do.
5cats said:
Oh, I wasn't even suggesting that that's how we'd go about it at all. It doesn't matter.
Your claim is that it's impossible to build a machine out of hardware OR software that is sentient. I disagree.
My line of reasoning goes as follows: Humans are machines. What would prevent other machines, even if they're designed by humans, to be sentient as well? Are you saying humans are special in some regard that can't be duplicated?
Whow! That was a long read (somehow I missed this thread, although it's exactly three of my favorite subjects - AI, chess, and civ together).
Of course, there were far too many not so recent ideas here - like the old "binary tree" analysis that, of course, is usually implemented only in enviroments where you have FULL KNOWLEDGE of the world: chess is such an enviroment, civ usually isn't, so you have to use something else. The other difference is the fact that the enviroment in civ is not fully deterministic - it's stochastic (you have odds, not certainties). That again makes extremely unsuitable the binary tree. Plus, I leave aside the FACT that even in chess computers do not really solve "to the end" each positions, but instead EVALUATE intermediate results (much more efficient).
If you want to look for a better "game example" to see what AI can and can't do, you should better look for other kinds of games. One more suitable example is backgammon, where you also have the "stochastic nature" of the dices. Another example could also be bridge, where you have the "not known world". But definitely chess isn't the appropriate pattern for civ (it's too different in two crucial aspects). If you look at AI achievements in such games, you will find examples like the remark of the World Champion in backgammon after a match against a program: he said that "it's STRATEGIC evaluation is far better than mine" (note that he is not talking about AI's CALCULATING ability). That means, there are statistical methods available that could create an AI that plays extremely strongly in civ - the main point is that AI must first use some method of statistical learning.
But even if that was done, with tremendous cost and effort, I really don't think it would be funny to play against the perfect AI in such a game - especially in levels with handicaps for the player, you could safely know that you would lose. Plus, that wouldn't fit easily with the "fixed rules" in civ AI, where each leader has a set of rules that determine whether he would usually build units or buildings or wonders. Stochastic variables + fixed rules + ability to be versatile aren't the best combination generally.
So, I wouldn't dream of a perfect AI opponent - but I don't doubt that you can construct an AI that plays extremely well (especially in domains like micromanagement, city specialization, etc., I don't think that any human could have a hope, when you have more than a few cities - we tend to be sloppy). Just make civ AI a bit more "unpredictable" (give him the right to do different things, so as not to know EXACTLY what AI will do in next turn, especially in wars) and it will be fine for me.
Of course, there were far too many not so recent ideas here - like the old "binary tree" analysis that, of course, is usually implemented only in enviroments where you have FULL KNOWLEDGE of the world: chess is such an enviroment, civ usually isn't, so you have to use something else. The other difference is the fact that the enviroment in civ is not fully deterministic - it's stochastic (you have odds, not certainties). That again makes extremely unsuitable the binary tree. Plus, I leave aside the FACT that even in chess computers do not really solve "to the end" each positions, but instead EVALUATE intermediate results (much more efficient).
If you want to look for a better "game example" to see what AI can and can't do, you should better look for other kinds of games. One more suitable example is backgammon, where you also have the "stochastic nature" of the dices. Another example could also be bridge, where you have the "not known world". But definitely chess isn't the appropriate pattern for civ (it's too different in two crucial aspects). If you look at AI achievements in such games, you will find examples like the remark of the World Champion in backgammon after a match against a program: he said that "it's STRATEGIC evaluation is far better than mine" (note that he is not talking about AI's CALCULATING ability). That means, there are statistical methods available that could create an AI that plays extremely strongly in civ - the main point is that AI must first use some method of statistical learning.
But even if that was done, with tremendous cost and effort, I really don't think it would be funny to play against the perfect AI in such a game - especially in levels with handicaps for the player, you could safely know that you would lose. Plus, that wouldn't fit easily with the "fixed rules" in civ AI, where each leader has a set of rules that determine whether he would usually build units or buildings or wonders. Stochastic variables + fixed rules + ability to be versatile aren't the best combination generally.
So, I wouldn't dream of a perfect AI opponent - but I don't doubt that you can construct an AI that plays extremely well (especially in domains like micromanagement, city specialization, etc., I don't think that any human could have a hope, when you have more than a few cities - we tend to be sloppy). Just make civ AI a bit more "unpredictable" (give him the right to do different things, so as not to know EXACTLY what AI will do in next turn, especially in wars) and it will be fine for me.
Zombie69
Emperor
- Joined
- Nov 22, 2005
- Messages
- 1,902
5cats said:
There's a lot of stuff in quantum mechanics that i don't agree with, and that i truly believe will be disproven in due time. You see, i believe in absolute determinism.
Einstein himself said that nothing can travel faster than light. Are you now telling us he said the opposite?
5cats said:
What's the clock in the box experiment?
5cats said:
Your opinion only. You probably also believe in the concept of soul. I don't.
warpus
In pork I trust
^^ Tachyons are theorized to move faster than light. I'm not sure what FTL travel has to do with civ and/or chess though.
dbergan
Warlord
First Questions on AI...
1) What is intelligence? The etymology of the word tells us inter-"between" + legere-"choose, pick out, read" inter-lege-nt-ia, literally "choosing between"
Therefore the essential idea behind intelligence is free will. Moreover, skill in using one's free will. What makes Kasparov intelligent and the cycle of the tides not? The fact that the tides are completely predictable, whereas Kasparov has the capacity to choose his own chess move. What makes Kasparov more intelligent (at chess) than David Bergan? The fact that his choices lead more often to victory.
2) Do computers have free will? Nope. They are completely deterministic. The closest they get is a mix of accurate calculations with a dash of randomness. Randomness in computer science is usually pseudo-randomness... the comp uses a function that gives us a distribution that simulates randomness, but the input and the output itself is deterministic.
More sophisticated computer random generators use things like the sound level (the third decimal point in the decible reading, for example) in the room so that it is not completely mathematically deterministic, but it is still deterministic. The function pulls all these variables in and gives you something that seems pretty close to a truly random distribution.
Of course we could say that dice and coin tosses are also deterministic. Randomness, like intelligence, is hard to pin down. But the notion is that modern physics does assert that some parts of quantum theory are truly random... Which is the meaning behind Einstein's quote that so many people use in their sig "God does not play dice with the universe." Einstein said that because he didn't think that there was a true randomness, but it looks like he was in error.
3) Do humans have free will? An ancient question. If free will doesn't exist, then neither does intelligence. If it does exist, is intelligence natural, supernatural, or both? If there's a supernatural element to intelligence, how does it interact with nature? (ie How does the soul command the body? Where is the connection? The pineal gland?)
On the other hand, if there isn't a supernatural element to intelligence, how do we explain awareness? (ie "It is impossible that our rational part should be other than spiritual; and if any one maintain that we are simply matter, this would far more exclude us from the knowledge of things, there being nothing so inconceivable as to say that matter knows itself. It is impossible to imagine how it should know itself."—Pascal, Pensées)
The questions only get more vexing the deeper you dig. What's funny to me is to hear people talk with certainty on the subject. You can state that you don't believe in souls... but be sure that you are clear that it is only a belief. There is no further evidence for one position or the other... you are merely stating that you are personally more comfortable with the idea that there are no souls, and therefore no free will, and probably nothing supernatural.
On the other hand, I prefer to think there are souls and free will because without them there is nothing more absurd than to thank someone for passing the mustard.
1) What is intelligence? The etymology of the word tells us inter-"between" + legere-"choose, pick out, read" inter-lege-nt-ia, literally "choosing between"
Therefore the essential idea behind intelligence is free will. Moreover, skill in using one's free will. What makes Kasparov intelligent and the cycle of the tides not? The fact that the tides are completely predictable, whereas Kasparov has the capacity to choose his own chess move. What makes Kasparov more intelligent (at chess) than David Bergan? The fact that his choices lead more often to victory.
2) Do computers have free will? Nope. They are completely deterministic. The closest they get is a mix of accurate calculations with a dash of randomness. Randomness in computer science is usually pseudo-randomness... the comp uses a function that gives us a distribution that simulates randomness, but the input and the output itself is deterministic.
More sophisticated computer random generators use things like the sound level (the third decimal point in the decible reading, for example) in the room so that it is not completely mathematically deterministic, but it is still deterministic. The function pulls all these variables in and gives you something that seems pretty close to a truly random distribution.
Of course we could say that dice and coin tosses are also deterministic. Randomness, like intelligence, is hard to pin down. But the notion is that modern physics does assert that some parts of quantum theory are truly random... Which is the meaning behind Einstein's quote that so many people use in their sig "God does not play dice with the universe." Einstein said that because he didn't think that there was a true randomness, but it looks like he was in error.
3) Do humans have free will? An ancient question. If free will doesn't exist, then neither does intelligence. If it does exist, is intelligence natural, supernatural, or both? If there's a supernatural element to intelligence, how does it interact with nature? (ie How does the soul command the body? Where is the connection? The pineal gland?)
On the other hand, if there isn't a supernatural element to intelligence, how do we explain awareness? (ie "It is impossible that our rational part should be other than spiritual; and if any one maintain that we are simply matter, this would far more exclude us from the knowledge of things, there being nothing so inconceivable as to say that matter knows itself. It is impossible to imagine how it should know itself."—Pascal, Pensées)
The questions only get more vexing the deeper you dig. What's funny to me is to hear people talk with certainty on the subject. You can state that you don't believe in souls... but be sure that you are clear that it is only a belief. There is no further evidence for one position or the other... you are merely stating that you are personally more comfortable with the idea that there are no souls, and therefore no free will, and probably nothing supernatural.
On the other hand, I prefer to think there are souls and free will because without them there is nothing more absurd than to thank someone for passing the mustard.
warpus
In pork I trust
dbergan said:
Perhaps we do and perhaps we don't. (have free will, that is - edit)
Does it matter?
We are sentient. We are machines. What would prevent us from constructing machines that are sentient as well?
The only thing I can think of is technological limitations.
If nature is able to construct a machine with free will, so can we, given enough technological advancement.
You can attempt to define intelligence all you want and go off on philisophical tangents about free will, but can you argue my point?
dbergan
Warlord
Zombie69 said:
How do you know? How do you explain awareness then?
dbergan
Warlord
warpus said:
If by machine, you mean deterministic (ie we have no free will), how do you know we are machines? And what is "sentient" if not awareness? So you still have to answer Pascal's challenge: "there being nothing so inconceivable as to say that matter knows itself."
- Status
Similar threads
- Sticky
