I meant to expound a little on this: > (Not to mention that some algorithms are more scalable that others, I want > to talk about that in a minute.)
In humans we often try to measure "intelligence" with tests and we call them IQ tests. It has been said that IQ tests actually only measure your ability to take IQ tests because they fall quite short of actually summing up intelligence to a single number. In fact, you can't. It's probably impossible to accurately quantify all the qualities of intelligence. Many people are genius's at some things and well below average in other things. Nevertheless, in each human is the machinery required for intelligence, a kind of computer we call the brain. We don't really understand very much about how it works, but it seems to consist of many key features that AI tries to emulate, pattern recognition, memory, ability to reason, etc. I would like to get the groups thoughts on this - because I'm on shaky ground here and don't claim any particular insights here - but I have a general theory about AI. It seems to me that you can view "intelligence" as a kind of physics. I would probably compare it to horsepower or "power" in physics. You can use whatever unit you choose, perhaps "watt" Work is not the same as power in physics. Work is the product of force and distance over which it moves in physical terms. In automotive terms it takes the same amount of "work" to move an equally heavy vehicle 100 miles or kilometers for instance. Of course I'm ignoring other physical factors such as air resistance in order to simplify. A scalable program can do any amount of work if you are willing to wait any amount of time. But some have much more horsepower than others. Mogo is a high horsepower engine. Brown is a seized up engine with so many inefficiencies that it works against itself. It can't really do any work. My basic idea here is that intelligence isn't static. Even in humans, as an approximation, it isn't about whether you can solve a problem or not, it's more about how long it takes you to solve the problem. We don't really think about it that way, but I believe it to be (more or less) true. With AI, computer memory is analogous to human memory. It's more like memoization in computer science. Once we learn something for the first time, we can use it over and over again throughout our lifetimes without having to rediscovery it. Of course the human brain is "hardwired" for many things it has been said. Much of what we need to survive and be intelligent, we don't have to learn - it was in us the day we were born. Sometimes, people who learn a lot of facts are considered intelligent. Even if they don't really understand or have much practical wisdom, we are impressed with someone who has so many facts stored in his brain and this probably should be considered one facet of intelligence. Computer Go programs have all these elements in them. They have memory, hard coded knowledge, reasoning ability (a life and death analysis can be considered a kind of reasoning ability as can an alpha/beta search) and so on. Sometimes we consider the ability to "figure something out" as being the main component of I.Q. as opposed to just "knowing the answer." And this is probably fair, because one is like having a fish and the other is like learning how to fish, a more useful skill in the long run. A scalable program has the ability to figure things out. A non-scalable program must be considered the type of AI that either "knows the answer" or doesn't. Sometimes we pretend there is no distinction because we are so time-conscious. We say that it doesn't matter if it can figure something out because we are too impatient to wait for the answer. We can measure the I.Q. of a GO program in a (very) rough way by calculating the ELO strength of the program and the amount of running time to produce this level of play. It's wrong to not consider time in this formula. In human I.Q. tests, the clock is part of the test and rightly so. Almost every problem in an I.Q. test is of the nature that you could figure out the answer eventually (if you are persistent and focused) but the clock holds you back. Time is an important variable in human intelligence, the accomplishments of the most brilliant scientists are often measured by the total body of knowledge they are able to contribute in a life-time as well as the quality of that knowledge. So my theory here is that A.I. is not a static quantity divorced from time but that time is a very important consideration. Every reasonably written chess program, for example, is equally strong if you don't time-constrain them. But only the best ones get considered as "strong" or "intelligent." The programs that play poorly are called "stupid" but they are only stupid because they are not efficient, not because they can't (eventually) figure out the right move. The "intelligent" or strong programs are like Ferrari's and the "stupid" ones with low I.Q. are like the old budget cars with 6 or 4 cylinder engines that had no passing power. But both can make a cross country trip equally well. As I mentioned previously, I did the I.Q. test with an early version of Lazarus and compared it to a stable version of GnuGo. Even though Lazarus had a higher ELO rating on CGOS, I concluded that GnuGo had a higher I.Q. The total amount of "work" that GnuGo can do is severely limited due to a very tiny gas tank. It has only enough fuel (energy) to do a quarter mile. So Lazarus will beat it in any long race. So if Lazarus is set to play at the same time rate as GnuGo, GnuGo will win. Even though GnuGo is not scalable, we can pretend that if it could continue to use it's "intellect" effectively at longer time controls, it would beat Lazarus at any level. I could give my program to Lukasz Lew for an engine overhaul and rebuild. He could boost the horsepower by making the engine parts more efficient. Even if he changed nothing in the program other than it's speed, it would be an I.Q. upgrade. It would be snappier, quicker to figure things out and better. I don't think you can say that Lazarus is really stronger than that particular version of GnuGo even though it rated higher on CGOS. I think it's more correct, at least in some sense, to view these as handicap games - even though the handicap is not explicitly defined. I can beat stronger players in chess if we make the appropriate time-clock adjustments - that doesn't imply I am "smarter" or have more chess I.Q. In this model, I claim that A.I. (or I.Q.) is not the total amount of "work" you can do (by some measure) but a function of the amount of work you can do in a given amount of time. In the field of A.I. research, much is being done today that was not possible yesterday. Neural Nets would not have been very feasible in the days where 8086 ruled. In fact, UCT and Monte Carlo for GO would have been dead-ends 20 years ago. I totally agree with those that say it's important to find "smarter" ways to do things but I also believe this is just another way to increase the amount of useful work that can be done in a given amount of time. A finely tuned racing engine isn't just about raw power, it's about efficiency too because efficiency give you more power. The general rule (in my opinion) is that playing strength will require a huge amount of "power" because that's what A.I. is. This in no way implies that it should not be "efficient" or that it should foolishly squander resources (as an internal combustion engine does.) Instead it should be as efficient as possible specifically so that it can do more work. And your not going to squeeze much water out of a rock. You're not going to get a free ride. You are not going to produce a strong program that doesn't do an enormous amount of work. Naturally, you want to do that work as efficiently as possible. The reason you want the work to be as efficient as possible is so that you can do even more work, not because you are seeking the holy grail of a program that plays like a master with a few lines of clever code and a constant time algorithm. - Don _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
