Some of you will know that Average is Over contains an extensive discussion of “freestyle chess,” where humans can use any and all tools available — most of all computers and computer programs — to play the best chess game possible. The book also notes that “man plus computer” is a stronger player than “computer alone,” at least provided the human knows what he is doing. You will find a similar claim from Brynjolfsson and McAfee.
Computer chess expert Kenneth W. Regan has compiled extensive data on this question, and you will see that a striking percentage of the best or most accurate chess games of all time have been played by man-machine pairs. Ken’s explanations are a bit dense for those who don’t already know chess, computer chess, Freestyle and its lingo, but yes that is what he finds, click on the links in his link for confirmation. In this list for instance the Freestyle teams do very very well.
Average is Over also raised the possibility that, fairly soon, the computer programs might be good enough that adding the human to the computer doesn’t bring any advantage. (That’s been the case in checkers for some while, as that game is fully solved.) I therefore was very interested in this discussion at RybkaForum suggesting that already might be the case, although only recently.
Think about why such a flip might be in the works, even though chess is far from fully solved. The “human plus computer” can add value to “the computer alone” in a few ways:
1. The human may in selective cases prune variations better than the computer alone, and thus improve where the computer searches for better moves and how the computer uses its time.
2. The human can see where different chess-playing programs disagree, and then ask the programs to look more closely at those variations, to get a leg up against the computer playing alone (of course this is a subset of #1). This is a biggie, and it is also a profound way of thinking about how humans will add insight to computer programs for a long time to come, usually overlooked by those who think all jobs will disappear.
3. The human may be better at time management, and can tell the program when to spend more or less time on a move. “Come on, Rybka, just recapture the damned knight!” Haven’t we all said that at some point or another? I’ve never regretted pressing the “Move Now” button on my program.
4. The human knows the “opening book” of the computer program he/she is playing against, and can prepare a trap in advance for the computer to walk into, although of course advanced programs can to some extent “randomize” at the opening level of the game.
Insofar as the above RybkaForum thread has a consensus, it is that most of these advantages have not gone away. But the “human plus computer” needs time to improve on the computer alone, and at sufficiently fast time controls the human attempts to improve on the computer may simply amount to noise or may even be harmful, given the possibility of human error. Some commentators suggest that at ninety minutes per game the humans are no longer adding value to the human-computer team, whereas they do add value when the time frame is say one day per move (“correspondence chess,” as it is called in this context.) Circa 2008, at ninety minutes per game, the best human-computer teams were better than the computer programs alone. But 2013 or 2014 may be another story. And clearly at, say, thirty or sixty seconds a game the human hasn’t been able to add value to the computer for some time now.
Note that as the computer programs get better, some of these potential listed advantages, such as #1, #3, and #4 become harder to exploit. #2 — seeing where different programs disagree — does not necessarily become harder to exploit for advantage, although the human (often, not always) has to look deeper and deeper to find serious disagreement among the best programs. Furthermore the ultimate human sense of “in the final analysis, which program to trust” is harder to intuit, the closer the different programs are to perfection. (In contrast, the human sense of which program to trust is more acute when different programs have more readily recognizable stylistic flaws, as was the case in the past: “Oh, Deep Blue doesn’t always understand blocked pawn formations very well.” Or “Fritz is better in the endgame.” And so on.)
These propositions all require more systematic testing, of course. In any case it is interesting to observe an approach to the flip point, where even the most talented humans move from being very real contributors to being strictly zero marginal product. Or negative marginal product, as the case may be.
And of course this has implications for more traditional labor markets as well. You might train to help a computer program read medical scans, and for thirteen years add real value with your intuition and your ability to revise the computer’s mistakes or at least to get the doctor to take a closer look. But it takes more and more time for you to improve on the computer each year. And then one day…poof! ZMP for you.
Addendum: Here is an article on computer dominance in rock-paper-scissors. This source claims freestyle does not beat the machine in poker.