Arimaa Forum - Print Page


    
      
        Arimaa Forum
        (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
      

        Arimaa >> General Discussion >> Why the challenge is by 2020?
        
(Message started by: asfaklive on Feb 20^th, 2010, 2:42am)

Title: Why the challenge is by 2020?
Post by asfaklive on Feb 20^th, 2010, 2:42am

I am not sure if I am right or not. I read the paper "When will computer hardware match the human brain?"

I found the line

"At the present rate, computers suitable for humanlike robots will appear in the 2020s"

Was this line a motivation for setting the challenge by 2020?

Regards.

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Feb 21^st, 2010, 2:15pm

Omar can answer this better than I can, but I believe he was not comparing the Arimaa Challenge to anything else. He has a goal of promoting AI. He doesn't want his challenge prize to be won by brute force alpha-beta searchers with hand-coded evaluation functions. But he believes that as computers get faster and faster, eventually someone will be able to win the Arimaa Challenge with a brute force alpha-beta searcher using a hand-coded evaluation function. The cutoff of 2020 is made in hope of paying out the prize money only if someone does it the "right way". If there were no cutoff, eventually someone would do it the "wrong way".

Title: Re: Why the challenge is by 2020?
Post by omar on Feb 21^st, 2010, 10:56pm

Yes, Karl's answer summarizes it really well.

Also you are right that I picked 2020 because that is when personal computers are expected to be as powerful as a human brain.

In 2002 when I was about to release Arimaa I was searching the net for Moores law and computing trends and I came across this article by Ray Kurzweil:
http://www.kurzweilai.net/meme/frame.html?main=/articles/art0134.html?

In this article he predicts that:
"supercomputers will achieve one human brain capacity by 2010, and personal computers will do so by around 2020".

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Feb 22^nd, 2010, 8:55am

on 02/21/10 at 22:56:06, omar wrote:

Also you are right that I picked 2020 because that is when personal computers are expected to be as powerful as a human brain.

I did not know that, Omar. You are even more under the influence of the futurists than I thought! :) In this context, I am eager for you to contribute to the State of the Challenge 2010 (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi?board=events;action=display;num=1262545661) thread.

Kurzweil will doubtless be wrong, and indeed only nine years after he wrote the article you linked there are already slight signs of his being wrong. However, this is not a knock on Kurzweil in particular. Making predictions is hard, particularly about the future. If I tried to predict the future twenty and fifty years out, I would do far worse than he does.

By 2010 we were supposed to have a supercomputer with the capacity of a human brain. It's 2010 now, so where is it? Would Kurzweil say that the capacity is there and we have just failed to program it properly? Yet in other sections, he makes the software advances sound every bit as inevitable as hardware advances, so he can't escape the absence of human-like AI by saying that society as a whole isn't properly focusing its energies on duplicating the human brain.

By 2020 we were supposed to have a $1000 computer with the capacity of a human. What if we don't? Most particularly, what if Moore's law holds until 2020, so we got the predicted exponential growth in computing power, but still no home computer that is a decent conversationalist? I suspect Kurzweil will have to say that he had no clue how flippin' amazing the human brain is, and that is why his predictions about producing its equal were so wrong.

But I tip my hat to him for this: "That is why people tend to overestimate what can be achieved in the short term (because we tend to leave out necessary details), but underestimate what can be achieved in the long term (because the exponential growth is ignored)."

He presents Moore's Law as a specific expression of a general historical inevitability, which seems absurd to me. I think that Moore's law will break down, which itself will bust Kurzweil's grand generalization. Computing power per cost will be an S-curve. But I have to concede that if I am wrong and Moore's Law continues to hold for the next fifty years, the future will be a lot closer to what Kurzweil imagines than to what I imagine.

No matter what the future holds, we will look back and say that what actually did happen must have happened. If Moore's Law breaks down, everyone will say how obvious it was that it couldn't continue forever, and how everyone who thought it would keep on going was a kook. If Moore's Law doesn't break down, everyone will say how Kurzweil's Singularity had to happen and anyone who didn't see it coming was an idiot.

My prediction (based on history) is this: The future will be something so unexpected to anyone, we will all look like idiots. :)

Title: Re: Why the challenge is by 2020?
Post by RonWeasley on Feb 22^nd, 2010, 9:47am

on 02/22/10 at 08:55:01, Fritzlein wrote:

My prediction (based on history) is this: The future will be something so unexpected to anyone, we will all look like idiots. :)

Perhaps this is what people mean when they tell me I'm way ahead of my time.

Title: Re: Why the challenge is by 2020?
Post by Janzert on Feb 22^nd, 2010, 10:56am

on 02/22/10 at 08:55:01, Fritzlein wrote:

By 2010 we were supposed to have a supercomputer with the capacity of a human brain. It's 2010 now, so where is it? Would Kurzweil say that the capacity is there and we have just failed to program it properly? Yet in other sections, he makes the software advances sound every bit as inevitable as hardware advances, so he can't escape the absence of human-like AI by saying that society as a whole isn't properly focusing its energies on duplicating the human brain.

Not that I agree with Kurzweil, but from what I can recall he has been pretty consistent in his date for human level AI as "by 2030". Actually it was the first "bet" placed at longbets.org (http://www.longbets.org/1) in 2002.

Janzert

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Feb 22^nd, 2010, 12:06pm

Thanks for that link, Janzert. To my ears Kurzweil makes a much more persuasive argument than Kapor. I'm not persuaded that emotion is harder than cognition, that intelligence can't exist without physical bodies, or that our subjective experience of consciousness is inherently not reproducible. I just don't get the stance that passing the Turing Test is impossible. I merely take issue with Kurzweil's calculations of how easy/inevitable it will be.

Kurzweil is more explicit here than in the previously linked article that the software is a more difficult challenge than the hardware, which frees him from having to explain why 2010 supercomputers or 2020 off-the-shelf computers are not as intelligent as a human being even if he maintains they have the "capacity" of a human brain. However, for consistency, he would have to claim that the software developed by 2029 could pass the Turing test while running on a stand-alone vintage 2020 computer, or on a 2010 supercomputer. (In this he would be just like Omar and the Arimaa Challenge; Omar believes that eventually software will be smart enough to run on a vintage 2002 computer and beat all humans.)

Kurzweil's confidence that we can get proper software comes from the feasibility of copying a human brain. That's an ingenious argument because it doesn't rely on human programmers to ever figure out the algorithms of intelligence. I would suspect that coding smart software is an easier task than copying a brain, but if I am right about that, Kurweil's argument is all the more logically compelling.

So, as a skeptic, I must argue both that the human brain is harder to copy than Kurzweil thinks, and that Moore's law will peter out. (If Moore's law doesn't peter out, then the difficulty of copying a brain is irrelevant, as we'll get there a few years later at worst.) Hmmm... maybe I'm skating on thin ice after all. ::) Kurzweil could be wrong about many of his super-optimistic claims and still win the bet.

I wonder what real-world events it would take to dent Kurzweil's optimism, and to make him fear that he was going to lose this long bet. His argument sure doesn't sound like the position of someone who thinks it is 50-50 or even 75-25 in his favor. He sounds 100% committed.

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Feb 22^nd, 2010, 12:20pm

on 02/21/10 at 22:56:06, omar wrote:

Also you are right that I picked 2020 because that is when personal computers are expected to be as powerful as a human brain.

It is ironic that Arimaa is more likely to make you rich and famous if you are wrong about AI.

Title: Re: Why the challenge is by 2020?
Post by Janzert on Feb 22^nd, 2010, 2:27pm

Of course these prediction dates have the rather, err interesting trend of following the Maes-Garreau Law (http://en.wikipedia.org/wiki/Maes-Garreau_Law). :)

Janzert

Title: Re: Why the challenge is by 2020?
Post by omar on Feb 25^th, 2010, 5:57pm

on 02/22/10 at 08:55:01, Fritzlein wrote:

I did not know that, Omar. You are even more under the influence of the futurists than I thought! :)

Yes, I've been for a long time :-)

It started back in 1993 when I was working at NASA and attended a symposium they organized where futurists were invited to come and share their views on the coming era of cyberspace.

http://www.nasa.gov/centers/glenn/news/pressrel/1993/93_17.html

This was the symposium where Vernor Vinge first proposed the concept of the singularity.

http://mindstalk.net/vinge/vinge-sing.html

I found the symposium quite fascinating although I thought that Vinge was a bit off his rocker to think that we would be able to create superhuman intelligence in just 30 years.

No doubt we would eventually have hardware that has the information processing capabilities comparable to the human brain, but that doesn't mean it will be as intelligent as humans. I've always thought the software problem was much harder.

Title: Re: Why the challenge is by 2020?
Post by omar on Feb 25^th, 2010, 6:33pm

I am considering making a prediction on longbets.org related to Arimaa. The prediction would be:

An Arimaa playing program that uses a tree search algorithm and a human generated evaluation function will not win the Arimaa challenge before 2020.

Would anyone like to take the other side?

Title: Re: Why the challenge is by 2020?
Post by 99of9 on Feb 25^th, 2010, 8:03pm

What is your definition of "uses a tree search"? For example, does bot_Rat's methodology count? how about mcTS?

Title: Re: Why the challenge is by 2020?
Post by omar on Feb 25^th, 2010, 9:56pm

Yes, the definitions would be a problem. I was originally thinking of say "... will not win without a significant break through", but then how do you define what is considered a break through. I thought stating it in terms of tree search would be more clear, but then what exactly is considered a tree search. I might have to think about this some more.

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Feb 26^th, 2010, 1:16pm

It would be interesting to get into definitions, Omar, because there is probably some bet that we could happily take opposite sides of. I'm growing confident that raw speed increases won't win the Challenge before the deadline, but I feel that some combination of speed increases, search enhancements, and evaluation enhancements added to what are in essence alpha-beta searchers, might be enough to win.

The bet that I want to make against you is that if the Arimaa Challenge is won by 2020, you will say, "Oh, that's just brute force, not intelligence." You will be disappointed. There is a long, embarrassing history of everything useful that computers can do being defined out of AI, and everything that remains inside of the definition of AI being useless, or not quite ready yet, or something that will have a societal impact ten years from now.

If the Arimaa Challenge is won, I will probably consider the insight that put computers over the top to be a breakthrough. MCTS for Go was a breakthrough, even if some people still don't think it is AI. Arimaa could have a similar "non-AI" breakthrough.

I'm pretty much over the issue of what is or isn't intelligence. Yes, I will be very interested in how the Arimaa Challenge is won, but I predict I will be bored by the debate of whether or not the winning method was intelligence. The fact that you want to rule out certain ways of winning as uninteresting, but you can't really define what you want to exclude, highlights our ignorance about what intelligence is beyond (A) it's what we do and (B) it's not what computers do, yet.

:)

Title: Re: Why the challenge is by 2020?
Post by omar on Mar 3^rd, 2010, 2:36pm

on 02/26/10 at 13:16:16, Fritzlein wrote:

The bet that I want to make against you is that if the Arimaa Challenge is won by 2020, you will say, "Oh, that's just brute force, not intelligence." You will be disappointed. There is a long, embarrassing history of everything useful that computers can do being defined out of AI, and everything that remains inside of the definition of AI being useless, or not quite ready yet, or something that will have a societal impact ten years from now.

I should accept this because I am guaranteed to win. No matter how it's won I just have to not be disappointed and say it was due to a breakthrough :-)

But seriously I don't know exactly what constitutes a breakthrough. Some breakthrough clearly stand out as breakthrough and others are more debatable. I also happen to think MCTS is a breakthrough though not as strong of one as say alpha-beta pruning. Maybe a breakthrough is characterized by a significant jump in performance. So much so that competitors cannot ignore it and have to embrace it to keep pace. That's what happened with MCTS.

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Mar 3^rd, 2010, 7:38pm

on 03/03/10 at 14:36:48, omar wrote:

I should accept this because I am guaranteed to win. No matter how it's won I just have to not be disappointed and say it was due to a breakthrough :-)

In fact, it is a metaphor for life. No matter what happens to you, just don't be disappointed with it; you will be happy! :)

Title: Re: Why the challenge is by 2020?
Post by RonWeasley on Mar 4^th, 2010, 5:30am

on 03/03/10 at 19:38:35, Fritzlein wrote:

In fact, it is a metaphor for life. No matter what happens to you, just don't be disappointed with it; you will be happy! :)

Except if you're being chased by giant spiders.

Title: Re: Why the challenge is by 2020?
Post by docreason on Mar 15^th, 2010, 9:00pm

on 02/21/10 at 22:56:06, omar wrote:

Well, we are now at 2010. I have no idea how close anyone is here to this. I would argue that no amount of speed added to a car will enable it to go underwater and function right (or into space). Methods, no matter how fast, can take you so far.

I also hope the IAGO leaderboard has some impact on this research also. I am hoping we eventually produce some generalists who are good around the board, and show raw ability to play all abstract strategy games well, with minor training. I hope for this so we can find a generalized path for the formating of strategic thinking.

Title: Re: Why the challenge is by 2020?
Post by Fritzlein on Mar 15^th, 2010, 9:33pm

on 03/15/10 at 21:00:37, docreason wrote:

I also hope the IAGO leaderboard has some impact on this research also. I am hoping we eventually produce some generalists who are good around the board, and show raw ability to play all abstract strategy games well, with minor training. I hope for this so we can find a generalized path for the formating of strategic thinking.

Although I am now an Arimaa specialist, I used to be a gaming generalist, so I have some intuitions about what it takes to be an all-rounder. My hunch is that a good generalist is not someone with generalized strategies or strategic pathways get used repeatedly, but rather someone who learns diverse specialized strategies quickly. The constraint on all players is time; if you want to play twenty games at a decent level you need to be able to quickly adapt to each. In particular, you need to be able to read and absorb whatever has been written about each. If there is no literature, you must be able gain insight from expert games.

In short, I expect that what will be special about the IAGO champion is not the way he thinks, but the way he learns.