When the Agents Come Marching In

Understanding how different systems of agents behave is the first step in garnering business value

In synchronized swim teams or marching bands, the individual performers respond only to cues from their neighbors as their role demands, but collectively they produce wonderful patterns for the audience to enjoy. If only business processes worked so smoothly and integrally. There is some hope that they can: Large systems of intelligent sub-units — or agents, as they are now called — when organized correctly can display highly intelligent and self-adaptive properties allowing modeling, forecasting, and even managing some of the complex interactions we find in business.

In fact, business rules, when set to trigger automatic actions, fit most definitions of the word “agent.” It makes sense, then, to ask yourself how you can make business rules work together on the micro level to create a spectacular result — a more intelligent enterprise — on the macro level.

I believe this is the first comparison study of different architectures of systems of agents (ASAs) and their implications for business applications. Each method for choreographing large sets of business rules in complex, interactive, and changing business contexts can be useful for very different classes of business problems.

Single-agent systems are cropping up all over the place and have become fairly common. For example, there is one on my computer that alerts me when airfare to my favorite cities falls by $50 or more. An example of a system that might generate and implement perhaps 20 to 200 independent, automated business action rules is Blue Martini’s e-commerce software. It can create decision trees on a collected base of electronic shopper data to create sets of rules that customize any company’s online store to the needs and characteristics of each shopper.

But what about systems of millions of agents? How can you organize them usefully, and what can you use them for? Most business processes are really based on thousands of small decisions that are integrated in some way, and that way is often less than obvious or optimal. By studying how different kinds of systems of agents behave, you can set up the backdrop for solving these more complex business problems.

Definitions of agents in the AI literature vary considerably but they all seem to roughly come down to anything that takes in information about its environment, interprets it, and then responds to it. These three processes are analogous to perception, decision making, and action.

This widely encompassing definition could apply to life forms and their components at many different scales: people, blood cells, brain neurons, ants, paramecia, and heliotropic plants, for example. The definition also includes artificial, intelligent units of differing complexity: robots; software objects; any sort of keyhole processes (see “Learning in Time,” Decision Support, March 20, 2000), including decision trees and artificial neural nets; units in a Markov Chain; and, important to this discussion, business rules — particularly when implemented for automatic action.

A business rule can be as simple as an automated trading rule (such as, if IBM’s stock rises above $100, buy) or an e-commerce Web screen display rule (such as, if the login is from a young male, show sports images). It can also involve complex multivariable triggers or multitiered decisions and actions.

ASAs could fall roughly into four categories: Muddy Waters would be what I’d call the ecological model with freeform interaction. Formal Affairs would be the one in which a set of structured variables or the problem’s formal structure organizes the agents. Agents at Outpost would have agents organized by inherent or discovered relations. And the last, Secret Agents, would be algorithms that can discover the system of agents and their interaction that most likely produced an observable phenomenon, such as the movement of a person or a market.

Here is an initial extrapolation of this classification of ASAs and some potential business applications of each.

Muddy Waters

The ecological model, with a main wellspring coming out of MIT from the Hive and the Amorphous Computing projects, is the most unstructured of the four ASAs and is inspired by the interaction of creatures in a pond. This modeling approach explores agents acting independently in a freeform medium. The research talks of “emergent properties” — the consequence of many independent autonomous units, each with its own intelligent response-structuring interaction, such as you might find in ecological systems. This philosophy is highly integrated with chaos theory, in which very simple functions when applied iteratively can lead to some of the most unexpected and complex structures, as with the production of the highly complex Mandelbrot set by the iteration of just one simple linear function.

The obvious business-decision applications that need exploring include modeling how many individual actors with different needs work together to produce different kinds of markets or group behaviors of customers, suppliers, and business competition. For example, there is tremendous potential for creating substantially more situation-dependent pricing models.

Formal Affairs

These are systems with a great many building blocks mapped directly to components of the problem and underlying database structures. This relationship makes the components and their relations easily readable, and makes global properties more easily traceable to component properties.

Because decision units in decision analysis software (such as Analytica from Lumina Decision Systems Inc.) fit the agent definition, the resulting formal diagrams mapping out business options can be seen as systems of agents. With these systems, for example, you can construct highly usable and interpretable Markov models for pricing derivative contracts or policy models for, say, studying the effects of regulations.

Most often, users of decision analysis software explicitly record business rules to use and base them on knowledge about the domain being modeled. This method contrasts with one in which the system learns decision-making rules automatically, as with decision trees and neural nets. This contrast between user-defined and data-defined agents is useful. Knowing when each is better to use is far from obvious, however.

Hierarchies of data-defined agents, such as the sort with which I experimented in “More for Less” (Decision Support, May 15, 2000) are useful in structured database information environments. In such an environment, the many small and localized decision processes that a business comprises could potentially be turned into automated decision processes and then into a globally optimized decision process.

An obvious application is to discover the real risk and return profile of combining all the incremental decision rules of qualitative money managers and then recombine them to create the best system for the business as a whole.

Traditional methods examine relations between historic track records, but this tack is actually quite unreliable because the analysis is highly colored by the time period chosen. The risk analysis in terms of composite rules is far more complete and can be truly optimized. Rules make explicit even the sometimes big risks a manager was just lucky enough to avoid in a given time period. The rocks are strewn with the wrecked galleons of investment heroes who had the hubris to proceed without knowing the true risks of their ventures, their client’s gold lost forever in deep waters.

Agents at Outposts

The next decision-support revolution, I expect, will include many distributed, electronic, intelligent analyzers working in distributed, unstructured information environments to help workers discover appropriate collaboration partners or illuminating information sources. In these situations the ASAs need to be structured by discovered relations in or between texts, networks, people, or data.

Autonomy Corp.’s knowledge management software has a lead in this area; it is able to match experts with other experts and experts with needed resources by creating agents that keep track of the “e-persona” or the proverbial hat a person is wearing at any given time.

Also fitting the Agents at Outposts category is Webmind, developed by Ben Goertzel of the Intelligenesis Corp. The method holds the potential, among many other things, to create specialized agents for each kind of medium — text, numeric, audio, and video — to allow the information in each to be equally sharable. His first system automatically determines and uses phrases in news stories indicative of changes in financial market conditions and combines this knowledge with more traditional numeric approaches for a substantially bigger forecast “bang.”

The Secret Agents

A whole new method coming out of Mitsubishi Labs in Cambridge, Mass. has exciting implications. This method, called “entropic estimation” lets you input a data set, say, the vicissitudes of a market or the movements of shoppers, and automatically produces a system of intelligent agents that generate these same kinds of observed interactive properties. This reverse engineering process reminds me of the Star Trek replicator, the thing that materializes a tasty, elaborate dinner from just raw atoms.

This approach allows explicit analysis of what are most likely to be the units and connections of the underlying dynamic process, such as the movement of a person in a store (real or virtual) or the flow of ideas in a discussion. But it also lets you simulate the system’s systematic reactions to ongoing variations of new conditions or to changes in structure. Because these systems are full of context-dependent inputs, you can see how a system generated in one context will react in another. This ability has tremendous potential for modeling ongoing customer dynamics — patterns of sequential buying, browsing, and eye-on-page movement, or customer retentiveness and satisfaction, for example — and forecasting how changes in promotions, store layout, page layout, automated Q&A, and customer service systems would affect these patterns.

As more and more advertising is competing for customer attention in the same densely packed virtual space, for example, having models of attention flow becomes a highly valuable marketing tool.

These basic four ASAs, though hardly exhaustive, are a good starting point for examining the implications of some of the new implementations of massive numbers of decision-making intelligent units in a larger business decision-making process.