Tip of the Week #17                     Tip Index

Go to the Prior Tip "Shareholders Aren't Everything"
Go to Next Tip "Planning for Crises in Project Management"
Return to MaxValue Home Page

Agent-Based Simulation

Simulation means different things. Most often, this refers to a model of a project or other system of interest. The ubiquitous computer spreadsheet is the tool of choice for most people preparing models, for example, a cashflow forecast model. Calculating the deterministic (single-point) model is one example of simulation.

More realism is achieved by recognizing uncertainties explicitly in the model. Probability distributions are a way to represent judgments about risks -- and to do the calculations. A popular way to solve for a stochastic (probabilistic) model is Monte Carlo simulation.

People have long modeled complex systems such as service facilities using discrete event simulation. Various players or objects in the system are modeled individually. Artificial intelligence techniques (especially data representation called "frames") led to attaching individual attributes to the objects in the system. A logical extension is to empower the objects with some autonomy. Sometimes, these autonomous objects are called agents. Perhaps the first agent-based modeling was Thomas Schelling's 1978 "Segregation Model" that demonstrateshow racially-tolerant neighborhoods can still become highly segregated.

Some problems are so complex as to defy simple description. There are uncertainties and non-linear behaviors. A trend or fad can feed upon itself (positive feedback). "Playing the game of life" by Rita Koselka (Forbes, April 7, 1997, pp. 100-108) is a fascinating article about the use of agent-based simulation to solve several interesting problems:

In the music album example, 50,000 imaginary people react to the introduction to a new recording. Parameters describe such things as how often the songs are played on radio stations and how virtual people react with one another. Let the computer crunch for two hours and the model produces some answers.

This brute-force approach provides a way of looking at problems that would otherwise not seem sufficiently world-like.

"Will these simulations make business absolutely predictable? Or enable governments to set policy so as to avoid undesirable unforeseen consequences? Of course not. But they certainly an make business planning more efficient and government policy less self-defeating. 'Only an idiot manages his business by spreadsheet, but only an idiot ignores insights you get from a spreadsheet or a simulation.'"

Forecasting Future Financial Crises

"Agents of Change" in Economic Focus,, The Economist, Jul 24, 2010, p. 76.

Most economists and their models failed to predict the 2008 financial meltdown and continuing problems. The gold standard for economic forecasting has been "dynamic stochastic general equilibrium" (DSGE) models. These models are and were used by the Federal Reserve and other central banks. They presume financial markets are efficient and the players are rational. The models perform well with minor perturbations yet utterly fail in crises.

In June, 2010, a workshop was held in Virginia sponsored by the National Science Foundation. The purpose was to
discuss the potential for agent-based models (ABMs) to predict and explain the recent financial events.  Participants included the Federal Reserve and other central banks. An ABM need not assume any sort of equilibrium. Individual agents can represent individual investors, lenders, and institutions. The behavior of each is described, such as valuing with fundamental analysis or technical analysis, degree of risk aversion, and adaptability to conditions and others' behavior.  Positive feedback can amplify affects, in contrast to mean-reversion (negative feedback) seen in traditional models. Positive feedback is evidenced by bubble generation and crashes. It may be someday possible to run ABMs in real-time to help manage the financial system, much like communities are doing with newly-implemented traffic-forecasting models.

"Complexity's Business Model" by Julie Wakefield, Scientific American, Jan. 2001, pp. 31-32, groups a number of technologies under "complexity theory": genetic algorithms, intelligent agents, neural networks, cellular automata, ant algorithms, and fuzzy systems. She describes the result of Southwest Airlines' agent-based model that reduced by 70% the amount of freight transfer. Agents in the simulation represented the behavior of every freight forwarder, ramp person, airplane and package. The simulation demonstrated that it was better to leave many packages on planes, even if that meant that the packages would not be going the most direct route to their destination.

Another article describes "Aspen" by Richard J. Pryor at Sandia National Laboratories. "A New Laboratory for Economists" by Otis Port, Business Week, Marcy 17, 1997. Used for simulating market behavior, the models can include 10,000 households plus 1,500 factories, stores, banks and government agencies. '"This is probably the best thing that's come along in a long time," says Lawrence R. Klein, a Nobel prize-winning economist and a pioneering economics modeler.'

"Virtual Management" in Business Week, Sept. 21, 1998, further discusses complexity theory (complicated behavior), adaptive agents, and cyber-bioengineering.   People are investigating large simulations of consumer and competitive behavior.   For example, department stores are interested in staffing levels and where to place cash registers and service desks.  "Instead of selling equations, they are selling pretty pictures.  Visualization is the marriage of mathematics and marketing (Michael Schrage)."

"Game Theory" by Allison Fass, Forbes, Nov. 14, 2005, describes several current uses of agent-based simulation for marketing studies.  Pepsi-Cola, for example, wants to optimize the location of vending machines in suburban offices. One simulation places a beverage machine in an office of 35 employees, and the simulation represents each person as an autonomous 'agent.'

The simulations are analogous to the popular SimCity and SimLife computer games.  It is "like a videogame without controlling the players."  They market models are forecastingówith Monte Carlo simulationóthe individual and collective group response to such variables as location, pricing, advertising, and purchase incentives.

— John Schuyler, April 1997 (revised 15-Sep-98, 21-Dec-00, and 2-Nov-05, 26-Jul-10)

Copyright © 1997-2010 by John R. Schuyler. All rights reserved. Permission to copy with reproduction of this notice.