2011/11/23 OrionWorks - Steven V Johnson <svj.orionwo...@gmail.com>: > A former girlfriend of mine, for which I'm happy to say I'm still on > good terms with, has a rated IQ of 150 or higher. Don't ever play a > game of GO with her. You will be humiliated. >
Interesting. Does she play Go in Internet? I was so humiliated that I today again lost to computer. It is so frustrating that not long time before, we could bark that there is no AI that can beat human IQ in Go, but then there came this Monte Carlo algorithm that guesses with small supercomputer (30 core cluster) some 30 million random playouts per second. And can thus beat most of the human club players by flipping the coin and picking up the moves that gives highest probability for winning. That is, it beats humans by using pure luck! However, although computers get stronger every year, Master level club players are still ahead of computers, and not to mentioning Grand Master level professional players. It will still take decades when computers can reach Grand Master level in Go. Some think that it is inevitable, others like me, think that there must be genuine breakthrough in AI-research before this can happen. And we have no way to predict when that will happen. see details, if interested from wikipædia: http://en.wikipedia.org/wiki/Computer_Go Anyway, Monte Carlo algorithm is very interesting approach to artificial intelligence, because it is not just brute force number crunching, but it tries to do a simulation that has as good as possible correspondence to reality. Also it is interesting, because it's approach to the game is very holistic. Usually we think that computers do in small and accurate details better than humans, but humans can beat computer in chess (while they still could beat computers back in 90's) using deep strategical understanding of the game. But with Monte Carlo approach it is other way around, that algorithm has very deep understanding of strategical principles and whole board thinking, but it often fails with small and isolated details. Because it just does not have enough calculation depth to read ahead the game with good enough accuracy. Therefore computer often blunders severely in tactical fights, because human reading ahead power is just superior, because humans can easily ignore irrelevant possibilities. Somehow, when I look computer to play Go, I think that it is rudimentary form of genuine artificial intelligence. It gives an impression that it would understand the idea of the game. I think that it is because, randomness is embedded into Monte Carlo algorithm so deep. Therefore it does not resemble what computers usually are associated with, i.e. deterministic deductive reasoning, but it tries to compensate the lack of certain knowledge of the game with creativity. Also there are more or less speculative ideas how human brain is functioning in neural net level. That also is based on huge amounts of random parallel branches (i.e. playouts), and then there is just picked branches that are associated to patterns of reality (with Monte Carlo there are searched patterns that gives highest winning probability). Therefore I would say that it is not far fetched that Monte Carlo algorithm's approach to the game of Go resembles human ability to recognize patterns. Because both are based on creating a simulation of reality that can predict future events. I do not think that if I made much sense, because these ideas are difficult to describe. But key idea is that I think that Monte Carlo algorithm is very good and promising approach to do genuinely intelligent AI. Also the key idea is that creativity emerges from randomness. –Jouni PS. There is curious Chinese professional Go teacher, who lives in Netherlands and gives Go lectures to Western audience. She more than often is mixing white and black stones while lecturing, although I think that she has more than 40 years experience of teaching and studying Go. So mixing should not be due to bad memory and inexperience what is the difference between black and white.
