How well does Advanced Go really work for reasonably good players (where I’m defining “reasonably good” to include 2k players, for this purpose)? What little knowledge I have of go software is well out of date, but I had the impression that unlike in chess, where computers were better than humans at tactics long before they were better overall, go programs aren’t terribly good at anything (on a full-sized board) unless one throws an outrageous amount of hardware at them.
What do you mean by “better at tactics long before they were better overall”? Getting the tactics right seems to be the point. Do you mean ‘good at seeking immediate goals but having a relatively poor lookup mechanism for evaluating possible future board positions’? Equivalently ‘Better at thinking n moves ahead but worse at guessing how good the n configurations will end up at n+5’.
Yes, that’s the kind of thing I mean, though perhaps for larger values of 5. It is customary for chess players to distinguish between tactics (stuff you can work out by searching) and strategy (stuff you can’t, where you play according to general principles / feel / high-level anticipation of what sort of thing will be happening several moves down the line).
Of course in the limit of outrageously effective searching strategy gets absorbed into tactics, but even in chess no player (human or computer) can look that far ahead. And in the not-exactly-limit of merely very effective searching, you can afford to be not quite so good at strategy if you can stomp your opponent tactically. This is generally how computers win.
The fact that computers and humans have distinctly different skill profiles is what makes “advanced chess” interesting: a hybrid with the strategic understanding of a good human player and the tactical monstrosity of a good computer player is very strong indeed.
(Having said which, I believe there’s some evidence that even a not-all-that-good human player armed with multiple computers running different programs can be scarily effective too.)
(Having said which, I believe there’s some evidence that even a not-all-that-good human player armed with multiple computers running different programs can be scarily effective too.)
If you’re thinking about the same thing I am, the player was “not-all-that-good” at chess, but knew a lot about chess programs and their different relative weaknesses and strengths.
Hypothetically, I wonder if that approach could be constructively imitated by a computer. A meta-chess program, dividing it’s computational resources between several subprograms, and combining their input to play better than the subprograms would if they had the full computational resources.
I think we are indeed thinking of the same instance. And yes, it would be interesting to try getting a computer to play that way.
Here’s a nice exploitation of a similar idea: The Fastest and Shortest Algorithm for All Well-Defined Problems; see also the discussion at Hacker News, where in particular you might want to read the comment from me that explains roughly what’s going on and the comment from Eliezer that explains one way in which Hutter’s description of his algorithm claims more than it really delivers. None the less, it’s a very neat idea.
How well does Advanced Go really work for reasonably good players (where I’m defining “reasonably good” to include 2k players, for this purpose)? What little knowledge I have of go software is well out of date, but I had the impression that unlike in chess, where computers were better than humans at tactics long before they were better overall, go programs aren’t terribly good at anything (on a full-sized board) unless one throws an outrageous amount of hardware at them.
What do you mean by “better at tactics long before they were better overall”? Getting the tactics right seems to be the point. Do you mean ‘good at seeking immediate goals but having a relatively poor lookup mechanism for evaluating possible future board positions’? Equivalently ‘Better at thinking n moves ahead but worse at guessing how good the n configurations will end up at n+5’.
Yes, that’s the kind of thing I mean, though perhaps for larger values of 5. It is customary for chess players to distinguish between tactics (stuff you can work out by searching) and strategy (stuff you can’t, where you play according to general principles / feel / high-level anticipation of what sort of thing will be happening several moves down the line).
Of course in the limit of outrageously effective searching strategy gets absorbed into tactics, but even in chess no player (human or computer) can look that far ahead. And in the not-exactly-limit of merely very effective searching, you can afford to be not quite so good at strategy if you can stomp your opponent tactically. This is generally how computers win.
The fact that computers and humans have distinctly different skill profiles is what makes “advanced chess” interesting: a hybrid with the strategic understanding of a good human player and the tactical monstrosity of a good computer player is very strong indeed.
(Having said which, I believe there’s some evidence that even a not-all-that-good human player armed with multiple computers running different programs can be scarily effective too.)
If you’re thinking about the same thing I am, the player was “not-all-that-good” at chess, but knew a lot about chess programs and their different relative weaknesses and strengths.
Hypothetically, I wonder if that approach could be constructively imitated by a computer. A meta-chess program, dividing it’s computational resources between several subprograms, and combining their input to play better than the subprograms would if they had the full computational resources.
I think we are indeed thinking of the same instance. And yes, it would be interesting to try getting a computer to play that way.
Here’s a nice exploitation of a similar idea: The Fastest and Shortest Algorithm for All Well-Defined Problems; see also the discussion at Hacker News, where in particular you might want to read the comment from me that explains roughly what’s going on and the comment from Eliezer that explains one way in which Hutter’s description of his algorithm claims more than it really delivers. None the less, it’s a very neat idea.