Good essay, although some misinformation
I liked the essay overall, but some of the information is incorrect or misleading: (the dotted lines bracket text cited from the essay)
-----------
Some of the factors responsible for the game tree's enormity are shown below. (A comparison to chess is shown in brackets)
...
- players can pass
-----------
This is misleading, since passes are rarely made during the game itself, but only at the end or close to it, since passing usually confers no advantage to a player during most of the game. Many variations appear in the rules on how to regulate passing.
Theoretically, if you include passing, another node should be added to every single branch of the chess and go minimax trees: a player could resign. This, however, is unnecessary, since it is rarely in a person's advantage to do so (except for psychological/etiquette reasons).
-----------
Go stones bore no relation to each other until the very end of the game. It was almost impossible to see which group an individual stone belonged to, and in fact players mentally placed stones in more than one group.
-----------
Not really. Often, it is quite simple for humans (sufficiently experienced in the game) to tell which stones are connected to which group. Many weaker players, in fact, sometimes make oversimplifications, considering stones to be part of a group (visually) when in fact they could be cut off with a few tactics. I do agree with the last statement: in certain cases, stones can be a part of more than one group, although more often this just connects the two groups into one.
-----------
One of the best Computer Go players is The Many Faces of Go which is rated on the Go handicap system at 6 kyu (the same as an average humn player with about 1 year experience).
-----------
This is a bit misleading. It is rare for a person to attain the rank of 6 kyu in one year (although it does happen enough to make it not extraordinary).
Also, there are many different ranking systems across the world. A rating of 6 kyu on IGS (the International Go Server, a rated, online playing network), for example, is equivalent to about 3 kyu in the American and Japanese Systems, and about 6 kyu on the Korean system, and about 4 or 5 kyu on the Chinese system. The lower the kyu number, the stronger the player is considered to be.
A rank beginner who has just learned the rules can be anywhere between 30 and 50 kyu, and after about 6 months of study, the person can usually attain 15-20 kyu, although results do vary widely.
Even though many programs ostentatiously brandish that they have achieved the ratings of 3 kyu or 6 kyu and such, in actual practice they have a much lower ranking. One test put Many Faces of Go to be at a bout 17 kyu on IGS, which translates to about 13 kyu in the Japanese and American systems. Players did not know they were playing a computer program (which does make a difference).
Once players realize they are playing a computer program, and once players have tried a few games with the program, often the players are able to quickly figure out the program's weaknesses and exploit them endlessly, resulting in very bad scores for the programs. This is not "cheating" on the part of the human players, however, and these types of games shouldn't be dismissed as skewed results. Playing strength is not determined by first time games but rather in a long term test of gameplay.
-----------
There is a database available to all known as the
Joseki database. This contains thousands of endgame
positions and optimal plays for each. Many human Go masters know most of them so to that end it doesnt seem like cheating to store them in a Computer player (except a computer has perfect recall).
-----------
Firstly, Joseki is a set of corner **opening** patterns, not endgame patterns, and it contains only what professional players consider as the best plays they can think of in and of themselves (without considering other parts of the board).
Secondly, I doubt that there are many people who know **most** of the variations. There are a few, to be sure, but only the most dedicated can know most. One can easily become a very strong player without knowing even a small fraction of the possible variations and patterns of Joseki. Most players know only a small handful.
Thirdly, what is "cheating"? AI is AI. No human is expected to work out the entire game of Go based on first principles, nor should a computer. It would a miraculous feat for any human to be able to compose all of the Joseki by theory alone; Joseki (which vaguely means "established stones") is just the collection of the experience of thousands of years of high level Go playing and studying. No chess player is expected to work out an entire opening book by theory, nor is any chess program.
-----------
Perhaps there will eventually be enough computing power to brute-force search the game tree of Go, but will there be a good enough evaluation function? I hope not because Go is providing an excellent domain for AI researchers and i believe it will continue to bring out new ideas and approaches to game AI.
-----------
How is the Minimax tree not part of AI? AI can mean many things. Minimax is, of course, a very crude AI structure, and it emulates the rough outline of the human player's thought process (consider possible moves, look ahead through, assume other player plays well, choose best result, play corresponding move). Still, it is within the realms of AI, since it does create a program that seems to be showing intelligence. Yes, this point can be disputed on and on, but nonetheless, the Minimax tree is a valid AI construct, if not a very elegant or expandable one.
Overall, I enjoyed the essay, but I think that people should be aware of the (albeit minor) mistakes/ambiguities in the essay.
|