I have been meaning for a while to point to this discussion by Cog at The Abstract Factory of "the wisdom of crowds" as a distributed computing problem, since (among other things) it says some stuff I've been meaning to say.
"The wisdom of crowds" is just another name for "the behavior of distributed algorithms".When you think about it, "Let's exploit the wisdom of crowds!" really means: "Let's set up a whole bunch of independently acting, loosely federated entities, each with an incomplete view of the system, and let's make them do some cognitive task." In other words, if a crowd ends up having any wisdom, it will have arrived at it through a distributed algorithm.
Why does this matter? Two reasons.
First, it de-mystifies the concept. "The wisdom of crowds" is a phrase precisely calibrated to mystify the thing it denotes....
By contrast, the phrase "the behavior of distributed algorithms" is a more forbidding thing, one that highlights a crucial fact: all systems for extracting knowledge from "crowds" are, in fact, intricate constructions that achieve their results through precise engineering of the rules governing the crowd.
This leads into my second point. Any computer scientist who has tangled with distributed systems knows that designing a distributed algorithm that actually does what you want it to do is extraordinarily tricky. On the other hand, it is really easy to design distributed algorithms that, for deviously subtle reasons, end up prone to behaviors like wildly unpredictable, bizarrely pathological oscillations, race conditions, deadlock, livelock, network floods, etc., etc., etc....
Naïvely lauding the alleged "wisdom of crowds" obscures the critical issue, which is the design of the distributed algorithm --- i.e., the social organization of the crowd. What are its mechanisms for passing information? For reaching consensus? Where are the possibilities for feedback loops? What happens in the obscure corner cases that result from the interactions of all its features? Etc., etc.
And, in fact, the social organization of "crowds" is perfectly capable of producing "wildly unpredictable, bizarrely pathological oscillations", among other things, a topic wonderfully surveyed in Christophe Chamley's Rational Herds: Economic Models of Social Learning. (Disclaimer: Christophe is a friend of the family.) Quite small tweaks in the social organization can spell the difference between the kind of social learning which gives everyone a warm, fuzzy glow (and near-optimal convergence rates), and stuff reminiscent of Charles Mackay, even when all the agents are rational utility maximizers making full use of available information. Somebody could probably get a thesis out of working out the connections between Rational Herds and, say, Nancy Lynch's Distributed Algorithms.
(To Surowiecki's credit, he does acknowledge various possible ways in which crowds can go wrong, including herding. To his discredit, he doesn't seem to adequately appreciate the importance of designing institutions to keep this from happening, and the difficulty of doing so, and the publicity material for his book certainly doesn't, as Cog quite rightly complains.)
Computer, social and cognitive scientists are not the only people who consider multi-agent systems; lately physicists have gotten in to the act. After all, we have some ideas of our own about how the interactions of large numbers of individuals generate large-scale patterns, so (as a number of people have explained to me), economics must be just another many-body problem. While I do think statistical mechanics could contribute to the study of collective cognition, and related areas of inquiry (see, e.g., here), I have to say that most econophysics is distinctly unimpressive. My a priori speculation, backed by nothing other than casual empiricism, was that this is because econophysics is driven by the same mechanisms as physicists' attempts to explain evolution as just another many body problem. (This should be strictly distinguished from using statistical mechanics to better understand actual biology or sociology.) Now Wolfgang Beirl threatens to complicate that picture with some actual facts. Blogging as The Statistical Mechanic, he sketches an intriguing paper that straddles the sociology of science and econometrics: "Cointegration of multi-agent research networks with financial markets and in particular the Nasdaq stock market bubble".
It will be based on a simple search of the word "market" in abstracts of papers stored at xxx.lanl.gov for the years 1998--2004 with the following results:
1998 25
1999 49
2000 70
2001 108
2002 94
2003 94
2004 79
If Wolfgang is ever bored enough to actually perform this study, I believe it will be the definitive analysis of the social basis of econophysics, and an important contribution to our knowledge of the social misconstruction of reality.
The Collective Use and Evolution of Concepts; The Dismal Science; Complexity; Learned Folly
Posted at December 22, 2004 23:28 | permanent link