Notes on Clark Chapter 9
("Minds and Markets")
Econ 308: Agent-Based Computational Economics

Last Updated: 16 June 2006
Latest Course Offering: Spring 2006

Course Instructor:
Professor Leigh Tesfatsion
tesfatsi AT

Syllabus for Econ 308

Basic References:

Rodney Brooks (1994), "Coherent Behavior from Many Adaptive Processes," In David Cliff (ed.), From Animals to Animats 3, The MIT Press, Cambridge, MA.

Shu-Heng Chen, Chung-Ching Tai, and Bin-Tzong Chie (2002a), "Individual Rationality as a Partial Impediment to Market Efficiency" (ps,1230K), (zip,266K), pages 1163-1166 in H. John Caulfield et al. (eds.), Proceedings of the Sixth Joint Conference on Information Sciences, JCIS/Association for Intelligent Machinery, ISBN: 0-9707890-1-7. NOTE: This is an abbreviated conference version of the following paper.

Shu-Heng Chen, Chung-Ching Tai, and Bin-Tzong Chie (2002b), "Individual Rationality as a Partial Impediment to Market Efficiency: Allocative Efficiency of Markets with Smart Traders" (ps,2045K), (zip,534K), Chapter 17 (pp. 355-375) in S.-H. Chen (ed.), Genetic Algorithms and Genetic Programming in Computational Finance, Kluwer Academic Publishers.

Andy Clark (1998), Being There: Putting Brain, Body, and World Together Again, The MIT Press, Cambridge, MA.

Dave Cliff and Janet Bruton (1997), "Zero is Not Enough: On the Lower Limit of Agent Intelligence for Continuous Double Auction Markets" (html,36pp), Technical Report No. HPL-97-141, Hewlett-Packard Laboratories, Bristol.

Herbert Gintis (2000), Game Theory Evolving: A Problem-Centered Introduction to Modeling Strategic Interaction, Princeton University Press, Princeton, N.J.

D. K. Gode and Shyam Sunder (1993), "Allocative Efficiency of Markets with Zero-Intelligence Traders," Journal of Political Economy 101 (1993), 119-137.

Edwin Hutchins and Brian Hazelhurt (1992), "Learning in the Cultural Process," pp. 689-706 in C. G. Langton, C. Taylor, J. D. Farmer, and S. Rasmussen (eds.), Artificial Life II, Addison-Wesley, Reading, MA, 1992.

Deddy Koesrindartoto (2001), "Discrete Double Auctions with Artificial Adaptive Agents: A Case Study of an Electricity Market Using a Double-Auction Simulator," Creative Component for a Master of Science Degree, Iowa State University , Ames, Iowa, June 14.

James Nicolaisen, Valentin Petrov, and Leigh Tesfatsion (2001), "Market Power and Efficiency in a Computational Electricity Market with Discriminatory Double-Auction Pricing" (pdf,162K), IEEE Transactions on Evolutionary Computation, Vol. 5, No. 5, October, 504-523.

Herbert Simon (1982), Models of Bounded Rationality, Volumes 1 and 2, The MIT Press, Cambridge, MA.

Leigh Tesfatsion (2002), "Agent-Based Computational Economics: Growing Economies from the Bottom Up" (pdf,216K), Artificial Life, Volume 8, No. 1, 55-82.

Basic Concepts

Key Issues

1. Redux from Chapter 4: Can the embodied cognition approach explain the most advanced and distinctive aspects of human thinking?

At the end of Chapter 4 (p. 82), Clark lists a number of major conceptual and methodological challenges posed by the embodied cognition approach:

Clark subsequently concludes (p. 83):

"The key to integrating the facts about advanced cognition with the vision of embodied active cognition lies, I shall suggest, in better understanding the roles of two very special external props or scaffolds: language and culture."

In Chapter 9, Clark returns to this theme.

2. Wild Brains, Scaffolded Minds? (Clark, Section 9.1, pp. 179-180)

Clark (p.179-180) argues that the extension of the embodied cognition framework to encompass more advanced cognition involves three aspects:

Clark continues (p. 180):

"The idea, in short, is that advanced cognition depends crucially on our ability to dissipate reasoning: to diffuse achieved knowledge and practical wisdom through complex social structures, and to reduce the loads on individual brains by locating these brains in complex webs of linguistic, social, political, and institutional constraints."

"Human brains... are not so different from the fragmented, special purpose, action-oriented organs of other animals and autonomous robots. But we excel in one crucial respect: we are masters at structuring our physical and social worlds so as to press complex coherent behaviors from these unruly resources. We use intelligence to structure our environment so that we can succeed with less intelligence."

3. Lost in the Supermarket (Clark, Section 9.2, pp. 181-184)

Clark (citing other researchers as well) points out that standard economic theory, with its substantive rationality postulates, seems to work best for highly scaffolded choice problems and to falter or fail as the degree of scaffolding declines.

More precisely, standard economic theory works best in the presence of constraining policies and institutional practices. According to Clark (p. 182):

"What is doing the work in such cases is not so much individual cognition as the larger social and institutional structures in which the individual is embedded. These structures have themselves evolved and prospered (in the cases where economic theory works) by promoting the selection of collective actions that do indeed maximize returns relative to a fixed set of goals."

Clark continues (p. 182):

"In the embrace of such powerful scaffolding, the particular theories and worldviews of individuals may at times make little impact on overall firm-level behavior. Where the external scaffolding of policies, infrastructure, and customs is strong and (importantly) is a result of competitive selection, the individual members are, in effect, interchangeable cogs in a larger machine."

Clark ends Section 9.2 (pp. 183-184) with a discussion of a well-known agent-based computational study by Gode and Sunder (1993). These authors show that high market efficiency is consistently obtained for a class of continuous double-auction experiments conducted with "zero intelligence" computational traders who make random bids and asks constrained only by budget constraints. Thus, the high market efficiency appears to be attributable to the special institutional form of the continuous double auction independent of the learning behavior of the individual auction participants.

4. Do double auctions really represent such strong scaffolding that learning is irrelevant?

As indicated above, an affirmative answer to this question is suggested by Gode and Sunder (1993) for continuous double auctions. However, subsequent findings strongly caution against over-generalizing the Gode and Sunder findings.

For example, Dave Cliff and Janet Bruton (1997) demonstrate that the high market efficiency observed by Gode and Sunder (1993) requires that the "true" aggregate demand and supply curves for buyers and sellers constructed on the basis of their reservation bid and ask prices be symmetric. More precisely, the slopes of these curves, although opposite in sign, must be of approximately equal magnitudes. In general, the mean transaction price resulting from the trading by Gode and Sunder's zero-intelligence traders can be reliably predicted from the particular probability density functions used by buyers and sellers to generate their bids and asks. In the absence of symmetry, this mean transaction price can differ significantly from the demand=supply "competitive equilibrium" price P* required for market efficiency.

Deddy Koesrindartoto (2001) shows that market efficiency can be seriously degraded in a discrete double auction with midpoint pricing (i.e., prices set at the midpoints of bid-ask spreads) if traders learn via a well-known individual reinforcement learning algorithm due to Alvin Roth and Ido Erev. The reason is that the latter algorithm fails to update choice probabilities in response to zero-profit outcomes. This can substantially degrade market efficiency in the context of a double auction because zero-profit outcomes are prevalent in the early stages of the auction when traders are undertaking price discovery and failure to match due to ask prices exceeding bid prices is common.

Nicolaisen, Petrov, and Tesfatsion (2001) show that high-market efficiency is obtained in the auction experiments run by Koesrindartoto (2001) if buyers and sellers instead learn via a modified version of the Roth-Erev individual reinforcement learning algorithm that corrects for the zero-profit updating problem. On the other hand, the authors show that market efficiency is seriously degraded if, despite the presence of revenue, cost, and capacity differences among buyers and among sellers, the buyer population and the seller population each attempt to learn "optimal" bid and ask prices by social mimicry using population-level genetic algorithms. The authors caution (p. 522): "While the discriminatory double auction may reliably deliver high market efficiency when buyers and sellers refrain from inappropriate learning behavior, it may not be robust against the active exercise of bad judgement."

NOTE: For a more detailed discussion of the zero-profit updating problem arising for the original Roth-Erev reinforcement learning algorithm, and the modification of this algorithm introduced by Nicolaisen et al., see Section 5 of L. Tesfatsion, "Notes on Learning"(pdf,157K) .

Finally, Chen, Tai, and Chie (2002a,b) show that too much "smartness" can degrade market efficiency in a discriminatory double auction with discriminatory midpoint pricing. (See the basic references listed above for pointers to zip and postcript versions of these papers). Under this pricing protocol, a different price is set for each matched buyer and seller at the midpoint of their bid-ask spread. The authors show that it can be rational for buyers to bid higher then their reservation bid prices and sellers to ask lower than their reservation ask prices if they expect that this risky price offer behavior will increase their chances of having their offers accepted while still resulting in a midpoint price that ensures them a profitable trade. The authors permit their traders to evolve these types of high-risk but potentially higher-return pricing stategies via genetic programming. Their experimental findings show that this results in a relatively unstable price and lower market efficiency than when budget constraints are tightly imposed on traders as in Gode and Sunder (1993).

Bottom Line: Studies subsequent to Gode and Sunder (1993) have shown that learning can matter substantially for market efficiency even in the presence of strong scaffolding such as double-auction protocols.

5. The Intelligent Office? (Clark, Section 9.3, pp. 184-186)

Clark asks (p. 184): "What kind of individual mind needs an external scaffold?

Clark notes that a vital role for external scaffolding is strongly predicted by research on individual cognition, beginning with the work of Herbert Simon (1982) on satisficing, proceeding through connectionist ideas (artificial neural networks and parallel processing, cf. Section 3), and down to present day work on embodied cognition.

He concludes (p. 186):

"Much of what goes on in the complex world of humans may thus, somewhat surprisingly, be understood as involving something rather akin to the `stigmergic algorithms' introduced in section 4.3. Stigmergy, recall, involves the use of external structures to control, prompt, and coordinate individual actions. Such external structures can themselves be acted upon and thus mold future behaviors in turn."

"And this, indeed, is just what we should expect once we recognize that the computational nature of individual cognition is not ideally suited to the negotiation of certain types of complex domains. In these cases, it would seem, we solve the problem (e.g., building a jumbo jet or running a country) only indirectly - by creating larger external structures, both physical and social, which can then prompt and coordinate a long sequence of individually tractable episodes of problem solving, preserving and transmitting partial solutions along the way."

6. Inside the Machine (Clark, Section 9.4, pp. 186-190)

Clark notes (p. 186-187):

"But as surely as these (large-scale scaffolds) inform and scaffold individual thought, they themselves are structured and informed by the communicative acts of individuals and by episodes of solitary problem solving. One crucial project for the cognitive sciences of the embodied mind is to begin the hard task of understanding and analyzing this complex reciprocal relationship - a daunting task that will require the use of simulations which operate at multiple time scales and levels of organization."

Clark sites seminal work by Hutchins and Hazelhurst (1991), who attempt to model the interplay of individual learning, cultural and artifactual evolution, and patterns of inter-group communication. The authors considers successive generations of agents modelled as artificial neural networks with simple fixed architectures (a few linked processing units).

Despite the absence of any change in their neural architectures, the agents gradually evolve better external cultural artifacts in the form of symbolic structures representing moon and tide states. These cultural artifacts increase the collective ability of the agents to predict an environmental regularity - the relation of moon phase to tide state - important for the acquisition of shellfish and other important food resources.

It is interesting to consider the relation of this work on the evolution of "cultural predictors" to the co-evolution of "individual predictors" by traders in the Santa Fe Artificial Stock Market Model studied in Section II of the course.

Indeed, although a number of ACE researchers are now studying the evolution of forecasting rules - in particular for financial markets - typically these rules take the form of privately owned tools applied by individual agents for individual advantage, not cultural artifacts for the collective benefit of society as a whole. To date, very few ACE researchers have attempted to incorporate cultural evolution considerations in their models. One interesting exception is Gintis (2000).

For further discussion related to this issue, see Tesfatsion (2002).

7. Designer Environments (Clark, Section 9.5, pp. 190-192)

Clark asks (p. 191): Without central control, what stops agent behaviors from becoming chaotic and self-defeating?

Brooks (1994) considers three sources of constraint:

Clark (p. 191) proposes that a fourth source of constraint be added this list:

Clark concludes:

"The successes of (standard economic theory) emerge, within this paradigm, as depending largely on the short-term dynamics of responses strongly determined by particular kinds of institutional or organizational structures: structures which have themselves evolved as a result of selective pressure to maximize rewards of a certain kind." (p. 191)

"If our achievements exceed those of our forebears, it isn't because our brains are any smarter than theirs. Rather, our brains are the cogs in larger social and cultural machines... This machinery is, quite literally, the persisting embodiment of the wealth of achieved knowledge. It is this leviathan of diffused reason that presses maximal benefits from our own simple efforts and is thus the primary vehicle of our distinctive cognitive success." (p. 192)

Copyright © 2006 Leigh Tesfatsion. All Rights Reserved.