Agent-Based Modeling Could Take the Toll out of Tolls

As someone who spent the better part of 15 years commuting to work from a place with no access to public transportation, I developed a strong dislike for toll booths. Not because of the cost – I am OK with the notion that my extensive use of the highway infrastructure should cost me more money. But what drives me crazy (pun intended) are the blatant inefficiencies of many toll booths, at least in the New England area where I am most familiar with them.

Complexity Science and ABM Approaches to Brand Value

As Apple’s stock passed $300 as few months ago, it triggered quite a reaction in me – and not just because I sold in the $100’s.  I wondered how the value of this brand could be growing so quickly and, more importantly, how I missed it.  After some consideration about the topic I realized that a new form of intangible asset valuation is possible – one that combines complexity science and agent based modeling.

Traditional measures of brand value have been built from three static and discrete components:

  • The financial forecast of the firm that owns the brand

How Complexity Science Can Empower Us

You sometimes come across people whose passion and purpose are contagious, people from whom you want to be infected. Two such people I know have embarked on two very different but inspiring journeys. I want to share a little bit of what I know about them and their empowerment projects, in the hope that their passion and purpose will infect you too.

Competence Without Comprehension

In a previous post I described the hunch engine, an exploration-support tool based on interactive genetic algorithms. There is an intriguing parallel I want to expose in more detail between biological evolution and decision making: search and evaluation in decision making are similar to variation and selection in evolutionary theory. Search is all about creating a variety of options and possible answers to a query; evaluation is the process through which some or none of the options are selected. Nature thus provides us with a powerful metaphor for decision making, and in that context genetic algorithms are decision-support tools. With interactive genetic algorithms, variation is performed by a non-human device while options generated by the device are evaluated by a human being.

Cyber Security: Human Behavior Matters

In an earlier post, our beloved Jim Fallows wrote briefly about a DoD-funded cyber-security initiative named SENDS, for Science-Enhanced Networked Domains and Secure Social Spaces. The overall objective of SENDS is to promote and begin to demonstrate the concept of a science of cyberspace with an initial focus on security. The vision for SENDS, developed by Carl Hunt, Richard Raines and Craig Harm, is one that embraces the richness, diversity and messiness of cyberspace. Central to their vision is the idea that the social, economic and behavioral aspects of cyberspace, which are largely missing from the general discourse on cyber security and are certainly under-funded and under-represented in government-sponsored programs, are at the core of what makes cyberspace the complex adaptive system that it is. An inclusive, multi-disciplinary, holistic approach that combines the technical and the behavioral is needed. Being a founding member of the SENDS initiative, I am definitely partial to its vision. The extent to which research and development in cyber security has been skewed toward “technical solutions” is mind-boggling. As an illustration, it seems surreal that in an otherwise excellent document, the authors of a 2009 manifesto from Sandia National Laboratories entitled “Complexity Science Challenges in Cybersecurity” have not dedicated a single line to human behavior. For example, their main M&S thrust is entitled: “Modeling the behavior of programs, machines, and networks”. No humans necessary – although I concur with the authors that there is a need for a new “cyber-calculus” – just the ability to frame concepts and issues in modern mathematical terms would be of enormous help. Or in a recent report by a group of DoD-funded physicists, you can read: On the positive side, the cyber-universe can be thought of as reduced to the 0s and 1s of binary data. Actions in this universe consist of sequences of changes to binary data, interleaved in time, and having some sort of locations in space. One can speculate as to why mathematics is so effective in explaining physics, but the cyber-world is inherently mathematical. But cyberspace, although it is the result of tremendous technological progress, is not just a piece of technology. It is both an enabler and an amplifier of human nature, eliciting new manifestations of human nature. It feeds (and in many ways feeds on) one of the most fundamental needs of human beings: communication. That it has become such an integral part of our lives in such a short time shows how deeply it resonates with our need to communicate and be connected. It should come as no surprise, therefore, that the multifaceted dynamics of cyberspace be so strongly influenced, even defined, by the behavior of its participants. According to Mark Graff of Lawrence Livermore National Laboratory: “[Cyberspace] gives individuals and small groups unprecedented reach to affect others; it makes physical distance much less of an insulating factor; confuses us about what is permanent, or public, or safe; and largely operates insensibly to us. We feel safer if important data is near us, or some place we know, or with someone we’ve met, but these comfort factors make no “Internet” sense and don’t scale to Internet dimensions either. In matters of risk assessment, we feel pretty safe from attacks originating “far away”; we also tend to ignore “low and slow”— or sporadic—attacks; random, “pointless” attacks (like from Internet worms) mostly tend to be low on our worry list, too.” No wonder that the intuition we have gained from the physical world over thousands of years of evolution leaves us ill-prepared to deal with the new geography of cyberspace. We can’t hope to acquire this new kind of intuition overnight. The bad news is that we suffer from severe limitations in our understanding of a critical component of our lives. The good news is that we are all subject to the same limitations – good news only if we can regain a competitive advantage in what has been a level playing field. Understanding our own behavior and that of our enemies becomes the most viable defense and the most potent weapon we can develop. Obviously it is essential to continue to improve the technical aspects of cyber security and significant investments need to be made to ensure continuous progress and to keep up with increasingly sophisticated enemies. But at the same time, human behavior is almost always the weakest link in security. The attacks on Google and other companies in China in 2009 were initiated through phishing –the underlying technical exploit is often trivial but social engineering is always the entry strategy. In the September/October 2010 issue of Foreign Affairs, Deputy Defense Secretary Lynn described the spread of a malicious worm on both classified and unclassified US Central Command systems, which started with the insertion of an infected USB key into a US military laptop. Apparently it took the Pentagon 14 months to clean things up. The worm would never have been able to infect any network without the help of someone like a malicious or clueless insider. On the flip side, the recent Stuxnet worm that damaged the Iranian uranium enrichment infrastructure seems to have used the same entry strategy of USB key insertion to get started; once in a system, it would use multiple exploits to spread itself. Example after example of intrusions and attacks point to the fact that human behavior is the enabling factor. In the case of the leaks of diplomatic cables to Wikileaks by Private Manning, human behavior is at the core: no technology solution would on its own prevent it. A small but growing community of scientists from academia and industry has emerged in the last few years. They need encouragement and support.

Predictive Modeling and Pricing in the Airline Industry

Complexity science manifests itself in all facets of life; one of my least favorites is the process I go through each time I have to arrange travel, something I unfortunately do often. The process would be sufficiently complex if I were simply dealing with the complexities of the trip itself, including times of meetings, times of flights, local transportation, weather and personal constraints. What makes the process absurdly complex is the arcane, short-sighted and consumer-unfriendly nature of airline pricing. Why should it be much more expensive to fly, say, Boston-Dallas round trip, than flying Boston-Dallas-Albuquerque round trip, even though the longer itinerary includes the shorter one? Why is it that sometimes the price of a ticket can literally double in the span of an hour while I am trying to resolve my travel constraints? Why is it so incredibly expensive and painful to change flights if, heaven forbid, my plans change? And why are refundable economy class tickets so expensive?

Time, Space and Interactions: The Case for Agent-Based Modeling (ABM)

When a customer makes a purchase or switch decision, it is often the result of a history. Impulse decisions, while they do happen, are the exception rather than the rule. That does not mean that most decisions are rational, simply that they cannot be explained by just looking at the time they happen. When a wireless phone customer decides to switch carriers, such a decision is the result of all the interactions and experiences this customer had with his carrier as well as with other sources of information. Failing to recognize the temporal dimension of decision making can lead to dramatic prediction errors. ABM, and only ABM, can explicitly deal with all aspects of time: learning, waiting, simmering, habituation, forgetting, interacting with other customers, etc. For example, in the casino industry, common wisdom holds that customers have a fixed budget and stop playing when their budget is exhausted. An ABM fed with real slot data from a loyalty card program showed that in reality customers stop playing when their total experience over time (TEOT) (a combination of the dynamics of their wins and losses weighted by demographic attributes and day of the week, and, yes, budget) reaches a threshold. TEOT is a much better predictor than budget or any combination of demographic attributes, which enables a major casino owner and operator to implement effective real-time marketing and promotional offers. Of course the dirty little secret is data and how to use it effectively to estimate complex time-dependent models of decision-making. When the data exists, in the absence of a coherent theoretical framework, not to mention theorems, one has to perform rigorous computational experiments based on statistical machine learning techniques. Another example is health insurance, where a customer’s demographic attributes are not sufficient to predict which plan he or she will select. Instead, the characteristics of each plan are viewed through a looking glass that puts more weight on certain characteristics as a function of the customer’s experience with his current plan, which is a combination of his and his family’s health in the past year and satisfaction or dissatisfaction with the health care afforded by the plan. Furthermore, if specific adverse health events have happened in the recent past, they strongly affect the way the possibility of catastrophic losses is perceived. By using an ABM that explicitly deals with the effects of experience and recency, prediction error could be reduced by an order of magnitude at Humana, a leading US health insurer. No amount of traditional econometric modeling with demographic attributes as explanatory variables would have been able to achieve this level of accuracy. In retail, the layout of a supermarket is known to be a key sales driver, yet shopper behavior is an emergent property with a strong spatio-temporal component that is never taken into account in traditional econometric modeling: while the trajectory of a shopper in a supermarket is influenced by the shopper’s shopping list, the trajectory in turn influences what the shopper buys beyond the shopping list. Through the use of a spatial behavioral model of shoppers in a supermarket, Pepsi was able to predict hot spots in any supermarket as a function of the supermarket’s layout and the demographic attributes of its shopper population. With the knowledge of hot spot locations, Pepsi could determine the best location not only for their products but also for promotional signs. Here again, the dirty little secret is data and how to use it. Not only did we have to develop special estimation techniques to infer trajectories, data collection itself was a challenge: shoppers were given “smart carts” equipped with tags for path tracking. Customers experience, learn, adapt, adjust. Their decisions are path dependent: in other words, decisions are dependent upon a contingent history. Existing statistical or econometric techniques do not deal satisfactorily with path dependence. When done properly (and that’s a big IF) ABM combines the statistical rigor of existing techniques with the ability to accurately model the temporal components of decision-making. As a result, not only does predictability go through the roof, the outputs of what-if scenarios also become more reliable because behavioral models are fundamentally causal rather than correlational. Knowing that two variables are correlated is good enough to predict the past, but robustly predicting the future requires understanding the underlying causal mechanisms of decision making.

index.php

Icosystem Corporation • 222 Third Street, Suite 0142, Cambridge, MA 02142 • Voice: (617) 520 1000

Terms of UsePrivacy Policy