The Wisdom of Critters and The Madness of Crowds

Can intelligent agents work together for better results, or is three a crowd?

The famous economist John Maynard Keynes made a fortune in the stock market following his own dictum, “think about what the average guy thinks the average guy will do.” In other words, in spite of his formal economic theory, he thought about the market system dynamics in terms of the response character of the players as they perceived and interpreted the other players. You could say he viewed the markets as a large system of agents perceiving each other mainly through the limited information source of price movement.

THE UNDISCOVERED COUNTRY

Agent Technology is Just Getting Started As I start out on the path of organizing ways of looking at large systems of agents, I am beginning to realize that we are entering vast new realms of thinking about computing. The vast array of kinds of intelligent units and the multiplicity of ways they can interact and communicate is so wide open. I feel like the ancient Greeks, who were aware that mathematics were important, but yet unaware of what was to come: the many disciplines and applications, such as group theory, analysis, differential geometry, genetics, demographics, space technologies, finance, and medicine to name just a few. Agents have the same potential to benefit and create disciplines of thought that we can’t even begin to foresee. The examples that I give, such as market analysis, traffic routing and scheduling, and other decision support tools are only a taste of things to come.

Agents are smart things that perceive, make decisions, and then act. This definition can include intelligent software objects, ants, people, business rules, and much more.

Large systems of agents, even when each individual agent has limited information or intelligence, can show remarkable total intelligence, as with ant colonies, bee hives, or general market pricing, or seemingly highly maladaptive behavior, as with lemmings or unsustainable market bubbles.

When do the combinations of many individuals or intelligent agents lead to systems of much greater intelligence, and when do they lead to the madness of crowds? The answer is in large part dependent on the system dynamics resulting from the properties and interactions of the parts.

Here I will give examples of some fantastically successful adaptive and fault-tolerant decision-support systems, modeled after the distributive computing power in ant colonies, and then I will focus on some of my rather simple, but remarkably suggestive, experiments simulating the ultimate swarm — the markets. Systems of agents can tell us a lot about market dynamics and how markets set prices.

I continue the discussion I started last time (“When the Agents Come Marching In,” July 17, 2000), when I introduced four different architectures for large systems of agents, by focusing on one that I call the Muddy Waters architecture.

One key issue that differentiates these architectures is the nature of the communication between the agents.

In the Muddy Waters architecture, just like creatures in a murky pond, communication and interaction is limited by some method in space and time. Bugs or agents interact only when they happen to come in physical, visual, or olfactory proximity or some link happens to occur between the bugs or agents, such as a long stick, a network link, or a dial-up to a broker.

The Hive project at MIT, for example, thinks in terms of people wearing agents and agents imbedded in everyday objects so that happenstance interactions can be more useful as they occur. For example, as you walk around your house at a given time of day, the right lighting, music, and beverage are there for your pleasure.

When a Software Bug Is a Good Thing

A lot of studies are being done on how insects, such as ants and bees that individually have simple behaviors, can generate complex social behaviors and on how to apply what is learned from these studies to create silicon creatures that can solve sometimes extremely difficult analytic problems.

Swarms of simulated ants laying down pheromone trails that slowly evaporate as they forage on a network of nodes have been used to rapidly find nearly perfect solutions to the nearly intractable traveling salesman problem — finding the shortest route, while visiting each of N cities only once.

This approach, with e-ant traffic a function of information traffic, is being researched to maximize the throughput of telephone switching networks, with tests suggesting that the ant-based methods are superior to Open Shortest Path First, the current Internet routing protocol. Potential applications are abundant in all sorts of optimal routing and scheduling problems, including transportation, stocks, products, production orders, and manufacturing.

These systems are highly fault tolerant and can dynamically reschedule when a breakdown occurs, just like real ant colonies when a food source runs out or a leaf bridge over a gully is blown away. A single problem node is rapidly routed around, and a problematic ant has a negligible effect on the whole.

A knowledge management application by Paul Kantor of Rutgers University lets colonies of Web users leave virtual pheromones at the sites that interest interests.

Clustering Critters

Some ant colonies sort their eggs, micro-larva or dead, by bringing similar to similar together one step at a time. Incremental clustering offers an algorithm that, unlike k-means, does not require an initial guess as to the right number of piles.

Both HNC Software Inc. and myself came up with similar algorithms without either of us knowing that ants have been doing the same thing, long before humans existed. My algorithm groups companies into industrial and sector categories by comparing their stock price profiles, and HNC’s Match Plus system applies to such diverse things as text clustering, automated answering of frequently asked questions, and detecting credit card fraud for real and virtual stores.

The Wisdom of Critters

Aristotle spoke of the wisdom of the masses in connection with democracy, not markets, but the issues are quite the same. Many individuals with limited information can produce something far more intelligent than any single individual — in the case of markets, setting prices, for example.

But there is a problem with traditional economics. It claims that a homogeneous lot of totally rational decision-makers with total information set a stable equilibrium price — four things that I rarely, if ever, see. Actually, markets fit the Muddy Waters architecture quite well.

All sorts of players might look at the market at some random time and decide to act for all sorts of reasons, with the vast majority of the information flow coming only from the narrow channel of prices.

Cognitive science research indicates that some of our reactions arise from emotional systems that are far less complex than we wish to imagine (we are agents made of agents), so the experiments I describe next may be more realistic than they appear at first glance.

In my little market critter experiment, I created thousands of simple agents, representing business or trading rules, which might check out the market at differing, random times, perceive the actions, think of others though the narrow channel of recent price changes, and react according to their character. Attempting to break apart complex behaviors into component behaviors is a way to see how the components and their variations affect and generate the whole.

First, I put into the pond a spectrum of risk avoiders that sold as soon as they made a range of tiny profits and a spectrum of those that saw a range of small dips as opportunities. Occasionally, I gave the waters a very weak stir (a factional point move, random in size and timing). I got processes typified by the green line in Figure 1, which does not look much like most market times series, because anything that might be either noise or information is quickly dampened.

FIGURE 1 Four different kinds of markets made by four different sets of trading critters.

In another pond, I put risk takers that jumped in the direction of movement (bought on up swings, sold short on down). I got processes like the purple line. This market overreacts, and tiny perturbations can be amplified many times over. You see large market swings some times representing more than 100 percent of value, even though very little information is available. This process is also not typical of most market times series, but it is part of the picture.

When I pulled down the wall between the two ponds and let these two differing ecosystems combine, I got a process typified by the yellow line. Now this process is starting to look like what you might expect in real pricing data for a company’s stock.

This approach immediately offers questions that with equilibrium theory you would not even think to ask: Does the market adapt so that the balance between risk takers and avoiders correctly displays the right balance between noise diminution and information amplification? Also, given any time period’s market conditions, what balance between risk takers and avoiders most fits its properties?

The Madness of Crowds

It is quite clear that, at times, risk taking is out of proportion with the realities of fundamental value, as with the Japanese real estate bubble or the recent Internet stock craze. (For more historical examples, see Charles MacKay’s Extraordinary Popular Delusions and the Madness of Crowds, 1841, Templeton Foundation Press, reprinted June 1995.)

What if I add another real world twist? The number of individuals willing to take risks grows as more people make money in the market, and when the market begins to fall, more buyers evaporate and more sellers come into the market. The blue line shows a typical result: a long bull market bubble, its rapid burst, and then an extended bear market.

It seems the markets show both the wisdom of the masses and the madness of crowds.

So why do the system dynamics for an ant colony work so well, while the system dynamics for markets work well most of the time, but not always? It is interesting to note that for real social insects, the unit of evolutionary survival is not the individual, but the hive or colony, and so genetic adaptation is in the direction of constantly improving the system’s performance. You can imagine the potential power of automated systems of agents where each individual agent evolves in order to increase the performance of the overall system.

However, buyers and sellers in markets are generally not interested in the performance of the whole. And this factor is why we give power to such entities as the Federal Reserve to cool the markets when steam bubbles seem to be appearing, a process justified by Keynesian economics. With a tightening or loosening of the interest rate screw, the balance between risk takers and avoiders hopefully shifts in the direction of more appropriate market pricing dynamics.

I have shown that large systems of even quite simple agents can produce both highly intelligent and adaptive decision-making systems and rather realistic-looking pricing dynamics that let us inspect and fine tune our reactions to hyper- or hypo-ventilated markets.

Keeping perspective, as with Keynes’s dictum, especially in times of rapid change, is what powerful decision making is all about, and studying how systems of rules interact, even in insect colonies, can be one way to get and maintain this valuable overview (besides offering tremendous potential for solving many sorts of other very difficult business problems). We may even gain four legs up on the competition.

Barry Grushkin (BLG23@cornell.edu) is Senior Researcher at the Machine Intelligence Company.

RESOURCES Ant-based optimization: iridia.ulb.ac.be/~mdorigo/ACO/ACO.html eHNC: www.ehnc.com Eric Bonabeau and Guy Théraulaz, “Swarm Smarts,” Scientific American, March 2000 Eric Bonabeau, Marco Dorigo, and Guy Théraulaz, Swarm Intelligence: From Natural to Artificial Systems, Oxford University Press, 1999 Hive at MIT Media Lab (Software Agents Group): lcs.www.media.mit.edu/groups/agents HNC Software Inc.: www.hnc.com Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, Prentice Hall, 1995