The Artificial Intelligence Thread

[bostjan]

and so Chess was solved quite early.

(citation needed)

Philosophical Magazine[/I]. 7. 41 (314).]With chess it is possible, in principle, to play a perfect game or construct a machine to do so as follows: One considers in a given position all possible moves, then all moves for the opponent, etc., to the end of the game (in each variation). The end must occur, by the rules of the games after a finite number of moves (remembering the 50 move drawing rule). Each of these variations ends in win, loss or draw. By working backward from the end one can determine whether there is a forced win, the position is a draw or is lost. It is easy to show, however, even with the high computing speed available in electronic calculators this computation is impractical. In typical chess positions there will be of the order of 30 legal moves. The number holds fairly constant until the game is nearly finished as shown ... by De Groot, who averaged the number of legal moves in a large number of master games. Thus a move for White and then one for Black gives about 103 possibilities. A typical game lasts about 40 moves to resignation of one party. This is conservative for our calculation since the machine would calculate out to checkmate, not resignation. However, even at this figure there will be 10^120 variations to be calculated from the initial position. A machine operating at the rate of one variation per micro-second would require over 10^90 years to calculate the first move!

It's an old article, but I have not seen nor heard of anyone solving Chess, other than some specific common endgame positions.

So...I'm going to challenge your claim that chess is a solved game. Post proof or else it hasn't happened yet.

 

[narad]

It's an old article, but I have not seen nor heard of anyone solving Chess, other than some specific common endgame positions.

So...I'm going to challenge your claim that chess is a solved game. Post proof or else it hasn't happened yet.

I didn't provide citation because I was referring not to some niche thing but to major events in AI, like Deep Blue defeating Kasparov in 1997. When you think of how that system worked then, plus Moore's Law, you have to really have a blind faith in humanity to think the best Chess player today is a human. Indeed, modern Chess computer efforts have ELOs higher than any human, and honestly these are not even reflecting the state-of-the-art paradigms in AI. I would hardly call them AI. And most often the challenge is not just making a powerful chess computer -- that is too easy -- but creating the most powerful chess computer in limited hardware configurations, like something that could run on a mobile phone or single intel cpu.

https://en.wikipedia.org/wiki/Human–computer_chess_matches

The only interesting research area now is not developing a computer that can beat the best human player, but one that learns to beat it in a novel way.

 

[narad]

Regarding your quote about complexity -- I didn't realize it wasn't my quote because of the formatting -- but it should be noted that the space for Go is far larger both in legal moves and in average depth, yet now no human can really stand any change against AlphaGo Zero. The article is phrased like the computer needs to exhaustively search positions, but of course this is not how anything works!

 

[bostjan]

Regarding your quote about complexity -- I didn't realize it wasn't my quote because of the formatting -- but it should be noted that the space for Go is far larger both in legal moves and in average depth, yet now no human can really stand any change against AlphaGo Zero. The article is phrased like the computer needs to exhaustively search positions, but of course this is not how anything works!

I think you don't understand how solving a game works. Perhaps we are talking about two completely different things because of the terminology.
A solved game is a game in which the best strategy is explicitly known. Meaning that we need to understand every possible strategy and be able to measure them all against each other. Neither go nor chess is solved.
In terms of being able to defeat a human player, a machine beating a human at any game doesn't impress me, but how the machine does it is interesting.
My takeaway from this is that, in your opinion, you don't think the application of this type of AI to chess is of any interest. But then again, you don't seem to have any interest in chess in general, nor enough interest in understanding the theories behind game design and game solutions to know the terminology commonly used, so maybe my opinion is that your opinion on this should leave me doing my own research into it. That's okay, though, not everyone can care enough about everything to engage a meaningful discussion, and it has nothing to do with your intelligence, just a mismatch of our interests.

 

[narad]

I think you don't understand how solving a game works. Perhaps we are talking about two completely different things because of the terminology.
A solved game is a game in which the best strategy is explicitly known. Meaning that we need to understand every possible strategy and be able to measure them all against each other. Neither go nor chess is solved.

I'm familiar with the terminology, but I think my sentence in context (or even the entire sentence, "Expert Chess players are great searchers, but machines are incomparably better, and so Chess was solved quite early.") is pretty clear that I'm referring to "getting a machine to beat a human at Chess" is solved. We are in the AI thread after all.

 

[bostjan]

Hmm. Beating someone at a game has never before been described as solving the game, IME. Well, I don't understand why you would use that term, knowing what it meant, to mean something so totally different, in that context, but it doesn't matter, now that we're clear.

Anyway, there had already been go programs that beat professional go players before AlphaGoZero, it's just that the new method without tactical training from a human is the first of it's kind to do so. I think it's completely relevant to Chess as well, but you seem to disagree as a matter of opinion.

Coming back to solved games, there would be no interest in any AI that can play a solved game, since you could easily hardcode the winning strategy. Even some unsolved games, like Poker, can be programmed with strategic algorithms to defeat human players in an arbitrarily long set of games.

 

[narad]

Hmm. Beating someone at a game has never before been described as solving the game, IME. Well, I don't understand why you would use that term, knowing what it meant, to mean something so totally different, in that context, but it doesn't matter, now that we're clear.

It's not a terminology thing -- the solving the game, in the game theoretical sense, does not follow in any way from the antecedent of that sentence. Whereas the research problem being solved does. This is known in pragmatics as "question under discussion."

I mean, this is the recurring theme of threads where we are talking, that you latch on to some formal meaning -- sometimes at the expense of it being relevant, sometimes simply through an understandable ambiguity in language -- and then harp on it. I don't mind clarifying, but I mind always having to justify the language in the post when it would be clear to most people what was meant.

Hmm
Anyway, there had already been go programs that beat professional go players before AlphaGoZero, it's just that the new method without tactical training from a human is the first of it's kind to do so.

There may be some confusion. Prior to AlphaGo (non-zero), no Go program had ever beaten a high-ranking Go player without handicap (sanctioned match). The handicaps make it hard to suss out, but they're significant -- 4+ stones. So to outdo Lee Sedol at 9P in an even match was huge, despite AlphaGo having tactical training from a human. This was the milestone, Nature paper (well tech. Nature paper was the euro Champion no where near as good as Lee), etc.

AlphaGo Zero, now without any human guidance, is significantly superior to AlphaGo.

To put it in broader context, the best pre-AlphaGo system was approximately 2100 ELO at best, or 7D, amateur dan. Pro is like 2700+. AlphaGo was 3100 ELO in the nature paper, 3700 vs. Lee Sedol, and 5100 ELO now.

I think it's completely relevant to Chess as well, but you seem to disagree as a matter of opinion.

For a computer (/established AI techniques), Go is the harder game. So while easier things may be interesting along some axis, definitely no one's going to see AlphaChess Zero and say, "Well now we've done it!" It'll just be like, yea, did that too. If you have an AI that can play Atari Montezuma's Revenge, you probably don't care much about Pong or Space Invaders, because there's just not as much going on there.

Like I'm sure there will be some interesting things to glean, but in the broader scope of AI there doesn't seem to be a killer research question particular to Chess, not in Go.

Coming back to solved games, there would be no interest in any AI that can play a solved game, since you could easily hardcode the winning strategy. Even some unsolved games, like Poker, can be programmed with strategic algorithms to defeat human players in an arbitrarily long set of games.

Agreed, but also why I wouldn't be referring to solved games in the earlier statement.

 

[bostjan]

It's not a terminology thing -- the solving the game, in the game theoretical sense, does not follow in any way from the antecedent of that sentence. Whereas the research problem being solved does. This is known in pragmatics as "question under discussion."

I mean, this is the recurring theme of threads where we are talking, that you latch on to some formal meaning -- sometimes at the expense of it being relevant, sometimes simply through an understandable ambiguity in language -- and then harp on it. I don't mind clarifying, but I mind always having to justify the language in the post when it would be clear to most people what was meant.



There may be some confusion. Prior to AlphaGo (non-zero), no Go program had ever beaten a high-ranking Go player without handicap (sanctioned match). The handicaps make it hard to suss out, but they're significant -- 4+ stones. So to outdo Lee Sedol at 9P in an even match was huge, despite AlphaGo having tactical training from a human. This was the milestone, Nature paper (well tech. Nature paper was the euro Champion no where near as good as Lee), etc.

AlphaGo Zero, now without any human guidance, is significantly superior to AlphaGo.

To put it in broader context, the best pre-AlphaGo system was approximately 2100 ELO at best, or 7D, amateur dan. Pro is like 2700+. AlphaGo was 3100 ELO in the nature paper, 3700 vs. Lee Sedol, and 5100 ELO now.



For a computer (/established AI techniques), Go is the harder game. So while easier things may be interesting along some axis, definitely no one's going to see AlphaChess Zero and say, "Well now we've done it!" It'll just be like, yea, did that too. If you have an AI that can play Atari Montezuma's Revenge, you probably don't care much about Pong or Space Invaders, because there's just not as much going on there.

Like I'm sure there will be some interesting things to glean, but in the broader scope of AI there doesn't seem to be a killer research question particular to Chess, not in Go.



Agreed, but also why I wouldn't be referring to solved games in the earlier statement.

Sorry that it seems like I harp on you using words that don't mean what you think they mean and having it confuse me. If it's a repeat offense, I assure you that it's not because I'm trying to single you out so much as it genuinely confuses me when you do it in each individual case. To my defense, you had not taken your first opportunity to clarify, but seriously, let's move on, now that I'm clear what you meant, even if I don't understand your choice of words.

I think you're missing the point with chess. I never said chess was the next natural step, only that chess was a game of similar difficulty with which I have more personal experience, therefore of interest to me. Go is not harder than chess, I think you're overstretching by beating that horse. Go and chess are equally difficult to win against a skilled opponent. Chess has a lot of opening book moves and endgame strategies used by humans, which are not at all understood to be optimal in any way other than how they work against other human players. When playing against a machine, historically, human players have tried to use unconventional opening moves in order to make the machine work harder by avoiding opening book assessments, and that has, in the past, generally benefited the human player. That means that a self-learning self-playing algorithm like what was used to make strong go play should be able to surpass other chess programs and help chess enthusiasts gain a deeper understanding of chess openings and maybe endgame strategies (which are already strongly dominated by machines).

-------

On a totally different thread of thought:

Discussion in the unpopular opinions thread about dance moves brought my mind back to a study in Nature about an AI (of some sort) algorithm that developed the "optimal sexy dance" based on learning from human preferences. With the whole Left Shark debacle a couple years ago and the recent developments in active real-time AI and also developments in manufacturing more realistic-looking androids, I was entertaining the thought of backup dancers being robots.

It might seem kind of silly as a use of all of this fancy technology, but, why not? If the AI is developed for other purposes, then it's application for more mundane tasks really won't be expensive. The groundwork is being done. These big concert tours for pop stars are really expensive...backup dancers are expensive...a robot dancer could potentially be more reliable and potentially less costly than a human, at some point in the near future, and, honestly, if the faces don't look super realistic from six feet away, who will notice during a concert, besides other performers? Then...the next thought is why use real musicians? If a robot could pantomime a dance routine, why not pantomime playing an instrument? The vocal tracks, as we know, are all prerecorded anyway, so, just play the damn mp3 and have robots do the entire visual show. Then the artist can just stay home.

-------------

Even less related: I have long been a participant in Amazon's Mechanical Turk: the artificial artificial intelligence site. Over the past couple of years, some of the tasks involving testing yourself against AI have gotten significantly more difficult. It's no wild guess that AI now is generally more advanced than it was even a couple of years ago. If you are interested, check it out. You can make a few pennies whilst you learn firsthand what some universities are working on with AI. One of the biggest things there right now is a project to scan youtube videos for offensive content. I think if you're interested in the nuts and bolts of practical AI, and you haven't checked it out yet, you'll have a blast there until it cuts you off.

 

[vansinn]

Fascinating discussion on chess and games.
@bostjan: Are you into game theory? I'm starting to feel a need for studying this area.

Unrelated to this specific game discussion, I believe there's a very good reason why AI won't beat human intelligence.
It's such a common mistake in mainstream science to believe in the usual ten dogma's of science, which almost exclusively states everything revolves around mechanics, and as such, also the mind, conscience, intelligence and more.
Essentially, that the brain is what all revolves around.

This type of thinking doesn't take into account that much of our intelligence sits in the body, in the shape of body memory.
Likewise, it doesn't factor-in that the heart is much more than a simple muscle pumping blood.
It's a fact that we have around 2500 nerve connections between heart and brain, which would be utterly unneeded, had the brain been the be-all of all things, and the heart mere a muscle.

How on Earth science will be able to digitize this.. well.. good luck with that.
Nope, AI will have it's multitude of applications, no doubt, but won't outsmart the human.

 

[bostjan]

I love game theory! I love the study of games as well, which is a sort of totally different thing.

Well, yeah, so there's a lot to what you're saying there, I think... a human being has an agenda, and that's always ultimately some form of self-preservation. Even through self-sacrifice, it's always in order to preserve something with which a person self-identifies, like their culture, family, nation, etc. The big huge gaping difference between human intelligence and AI is that the AI is more flexible with its ultimate goals, which will probably be to preserve the "self" with which its creator identifies.

In terms of "outsmarting," it's all a matter of how well AI can do something versus a human. AI can become a better go player than a human. AI will ostensibly become better at driving a car than a human. I think the averages on these things are just going to be so. There are psychologically interesting bits of those two claims, though:

AI becoming better at playing go than a human means that AI can be programmed to do something profound. At a high level, this is itself profound. At a low level, though, I mean, go is a game created for the purpose of stimulating the human mind. We don't need AI that can beat us at go. An AI that dominates the planet at playing go benefits neither the AI nor the humans of the world. But the fact that this exists means that there can be other applications. But, coming back to low-level implications, it also might be exciting for go players, as they will have something new to study in order to challenge them. Back to the high level, though, the fact that AI discovered it's own winning go strategy that humans had not yet reached implies that AI might have a broader reach into the field of new discoveries.

AI driving a car is a much more directly beneficial thing. Fewer car crashes would mean less mechanical attrition of resources allocated to repairs and it certainly would lead to fewer injuries of humans. But that AI serves a specific purpose and is something we would likely grow to rely upon in time.

I think the ultimate difference is that, at least up to this point in time, AI is still a tool for humanity. Does this mean that a situation like Maximum Overdrive or The Terminator could never happen? No, but IMO we're a significant distance away from that, still. It's not like AI made to drive a car or beat people at go is going to suddenly develop interest in enslaving humanity. On the other hand, once AI is prevalent, somebody will probably make an AI that could develop into something horrifying.

But like I said in another thread, the threat of that is real, but far less present than the threat of a superdisease epidemic or of a terrorist group obtaining a nuclear warhead, and causing significant damage to humanity as a whole. Those problems are a lot closer to home at the moment.

 

[StevenC]

Fascinating discussion on chess and games.
@bostjan: Are you into game theory? I'm starting to feel a need for studying this area.

Unrelated to this specific game discussion, I believe there's a very good reason why AI won't beat human intelligence.
It's such a common mistake in mainstream science to believe in the usual ten dogma's of science, which almost exclusively states everything revolves around mechanics, and as such, also the mind, conscience, intelligence and more.
Essentially, that the brain is what all revolves around.

This type of thinking doesn't take into account that much of our intelligence sits in the body, in the shape of body memory.
Likewise, it doesn't factor-in that the heart is much more than a simple muscle pumping blood.
It's a fact that we have around 2500 nerve connections between heart and brain, which would be utterly unneeded, had the brain been the be-all of all things, and the heart mere a muscle.

How on Earth science will be able to digitize this.. well.. good luck with that.
Nope, AI will have it's multitude of applications, no doubt, but won't outsmart the human.

It's 100% testable that the heart doesn't have any cognitive ability. But that's pretty sciencey so feel free to discount it.

 

[narad]

Sorry that it seems like I harp on you using words that don't mean what you think they mean and having it confuse me.

Again, I know the meanings of the words...

I think you're missing the point with chess. I never said chess was the next natural step, only that chess was a game of similar difficulty with which I have more personal experience, therefore of interest to me. Go is not harder than chess, I think you're overstretching by beating that horse. Go and chess are equally difficult to win against a skilled opponent.

Of course, but it's easier for a computer to be that skilled opponent by virtue of what computers are good at, and the inherent properties of Chess. If for every move in a game, any game, I could zoom into the future and find out the outcome of this move played out in totality, then I would never lose. Computers can exploit this to a far greater extent than humans, so if your game has a smaller branching factor and fewer states to unroll before reaching end-game, the greater the extent they can approximate this "oracle" strategy, and the better the computer will do. But this shouldn't feel like intelligence to anyone.

Now comparing chess to go, go has an enormously larger search space, which means it can rely less on computer-like talents and needs something more human-like.

To state this as succinctly as possible: we're able to learn interesting lessons about AI, and about go as a game, because AlphaGo has a special talent: the ability to look at a single state of the board and estimate who will win and their degree of advantage. It is not always easy to look at a game of Chess and determine who is in the lead, but far, faaar easier to do than Go, and combining that with lookahead gets you much closer to a useful system than in Go.

So on one side of the spectrum you have games you can completely unroll, score, and play an optimal strategy. This requires zero intelligence. On the other, one with a massive amount of moves, and where lookahead is very difficult, and therefore one must rely on a more intuitive understanding of what parts of the board contain the potential for advantages. This requires something closer to what we would call intelligence. Wherever go and chess fall on that spectrum, go falls a little bit more towards the latter side.

Naturally there are lessons learnable from AI in either game, that are useful for furthering our understanding of game strategies.

 

[bostjan]

I guarantee that a computer cannot zoom ahead an arbitrary number of moves in a game of chess to see all possible outcomes.

Naturally there are lessons learnable from AI in either game, that are useful for furthering our understanding of game strategies.

That's really all the gist of my comment was.

And one more thing on this topic we seem to get stuck on:

So on one side of the spectrum you have games you can completely unroll, score, and play an optimal strategy. This requires zero intelligence. On the other, one with a massive amount of moves, and where lookahead is very difficult, and therefore one must rely on a more intuitive understanding of what parts of the board contain the potential for advantages. This requires something closer to what we would call intelligence. Wherever go and chess fall on that spectrum, go falls a little bit more towards the latter side.

I don't care how massive the amount of moves is to look ahead, once you've done the entire game from start to finish with the objective best strategy and proven there is no strategy that beats it, you're finished with that game. It could be tic-tac-toe with a small enough number of games that you could work it out on paper with a pencil, or it could be draughts, which requires computer assistance to handle all of the data - but the end result will forever be all there is, and then the topic of that game's optimal strategy is done.

Whether it's chess or go or whatever, if the game is not solved, there is still some interest in it remaining. AlphaGoZero makes a huge leap toward potentially solving go, as it's come up with new strategies on its own.

So, @narad , what are your thoughts on the other topics I brought up? Have you ever participated in Mechanical Turk?

 

[narad]

So, @narad , what are your thoughts on the other topics I brought up? Have you ever participated in Mechanical Turk?

I'm familiar with it but haven't participated, so can't really comment.

It is interesting though that I saw some new machine translation system that was blowing away Google MT results. Turns out the company doing it, their main product was like a search engine for translators to go find example sentences. Perfect data. Then they just reimplemented what is basically Google's in-house system and trained it on way more data, and it was producing significantly better translations.

One could also turk that out, so there are these really game-changing advancements coming not from the tech, but just from ultimately someone paying people to create that data. It never goes away, so we just keep making incremental progress towards universal translation even without pushing the science.

 

[vansinn]

It's 100% testable that the heart doesn't have any cognitive ability. But that's pretty sciencey so feel free to discount it.

I weren't talking about cognitive powers in the heart, but our emotion set, which is a driving force to us.
Think of terms like, "a sting in the heart", "heart broken", "a loving heart", "two hearts falling in love", and so forth..

Try investigating Gregg Braden, and the documented heart-field, which is measurable (sorry about no providing links; I'm quite busy ATM).
We communicate by babbling off our mouth piece, as well as via pheromones, as well as mentally, that is via our mental capabilities, as well as from heart to heart, which were my reason for referring the 2500 connections between heart and brain.
And I didn't even mention our use of the pineal gland.. (for those who still haven't been [sodium fluorite] poisoned enough to have ended up with a partially defunct pineal).

But most science will not accept this, which was part of my point, another point being that AI won't have neither the heart field nor mental powers, and likely nor a pineal gland.

 

[narad]

But most science will not accept this, which was part of my point, another point being that AI won't have neither the heart field nor mental powers, and likely nor a pineal gland.

Science doesn't actively "accept" or "deny" things by some arbitrary value judgement. If the heart was the emotional center of the body then we might predict that a heart transplant would drastically alter someone's emotions. Because we find that this is not the case, a major prediction of your theory does not hold, and casts doubt on there being any emotional processing taking place in the heart. In contrast, even relatively minor brain injuries have been known to drastically alter people's behavior, often along emotional axes.

This is absolutely the nicest reply I'm capable of on this subject.

with a partially defunct pineal

Hey, who have you been talking to? I said I was stressed out at the time and that's never happened to me before!

 

[StevenC]

I weren't talking about cognitive powers in the heart, but our emotion set, which is a driving force to us.
Think of terms like, "a sting in the heart", "heart broken", "a loving heart", "two hearts falling in love", and so forth..

Today's Word of the Day:
Metaphor: a figure of speech that refers, for rhetorical reasons, to one thing by mentioning another thing

 

[vansinn]

No, I'm not referring to one thing while mentioning another.
I'm referring to which mechanisms constitutes the heart, and the huge number of connections between heart and brain, while comparing this to AI, which I can't see will have such.

I don't understand the rationale in quoting words of the day as a make-fun mechanism, while seemingly not looking into the mechanisms I refer.
Rather, it seems more like a continued wish to look at AI as something mechanical, which I likewise referred, as the limiting dogmatic pillars of science.

 

[StevenC]

No, I'm not referring to one thing while mentioning another.
I'm referring to which mechanisms constitutes the heart, and the huge number of connections between heart and brain, while comparing this to AI, which I can't see will have such.

I don't understand the rationale in quoting words of the day as a make-fun mechanism, while seemingly not looking into the mechanisms I refer.
Rather, it seems more like a continued wish to look at AI as something mechanical, which I likewise referred, as the limiting dogmatic pillars of science.

You can call it dogmatic, or you can list some of the great scientist through history and we can talk about how wrong they were about stuff. It's weird how most would call Isaac Newton one of the greatest scientists ever, but equally he was bad at calculus and his physics got trampled by Einstein.

Dogmatic science is bad science. Science is measured by irrefutability, and there's a good reason why theorems don't exist outside of mathematics.

You can't talk about the 2500 connections between the brain and heart and then cry when someone figures out they don't work the way you think they do. That's dogmatic.

We can engage rationally on the subject, or you can take your bad breath somewhere else (assuming you meant sodium fluoride).

--

Back to AI: @narad how are things going with AlphaGoSoccer?

 

[narad]

Back to AI: @narad how are things going with AlphaGoSoccer?

Well there's thousands of connections between the brain and feet, hence phrases like, "kicked the bucket", "get your sea legs", "a leg up [in this world]", and "pulling my leg". So it must be a long ways off.