Agent Based Modeling

In the past two weeks I’ve been to two very different conferences: The Santa Fe Institute’s annual Business Network Symposium in New Mexico and the Canadian Pension Fund’s Innovation Conference in Bermuda.  The same important topic was discussed at both conferences: Agent Based Modeling.  At Santa Fe, I did the listening.  In Bermuda, I did the presenting.

The traditional way to model economics or markets is called “top down:” you assume a a distribution for a stock’s returns or you assume a certain shape of the interest rate curve or something similar.  From there, you attempt to calculate things like the probability of portfolio loss or the long-term value of bonds.  In essence, you start with a high level assumption of how the world works and you then calculate determine how something of interest behaves.

Agent Based Models are very different: they don’t start with assumptions of how the world works.  Instead, they start with the players who interact – the “agents.”    They then incorporate the individual agents’ observed behavior in a given circumstance and they then let those agents interact.   They calibrate those “near-neighbor” interactions by changing them until the macro behavior of the ensemble resembles what is observed in reality.  In my opinion, it’s a much more natural modeling approach than the traditional mathematical top-down approach.

For example, suppose you want to model traffic flow on the highway.  Each car would be an agent.  Based on observed behaviors, the modeler would assign speed preferences according to some distribution and “rules” of how to behave.  Some drivers always want to keep a minimum distance from the car in front.  Others drive very close together.  Some drivers pass only whenever they come within 2 car-lengths of the car in front of them, and so on.  You can also include things like the effects of slowing down when there’s an accident on the other side of the barrier.  The point is that each agent has a set of rules which can be as simple or as complicated as you like.  You then let these agents loose in the virtual world you’ve created as see what happens.  Los Alamos Labs has done a lot of research on this exact problem and has helped cities like Portland, Oregon design better road.  One surprising conclusion: it’s not always better to add a lane to a highway or offramp as it can lead to people changing lanes too often, which disrupts traffic flow.

Another example of ABM’s use is in designing building exits — especially important during emergencies.  In this case, the agents are humans who are running for the door.  What type of exit structure gets the people out fastest and safest?  It turns out to be a very counterintuitive answer that has been shown to work in practice, all based on agent based models: put a column a few feet from the middle of the double door exit.  This column, which forces people to decide far away from the exit which side they will go to, actually speeds up everyone’s departure.  Hmmm.  Seems like these models are pretty useful!

Another example that I find personally important is how to most efficiently board an airplane.  ABMs have shown that filling a plane “from the rear” is actually less efficient than just letting everyone on in random order!  All those hours of my life wasted waiting to get on a plane and we could have been faster if they just let us all on at once!  But it turns out that an even more efficient method is boarding by odd/even window seats on alternate sides of the plane (first, odd rows on one side and even on the other.  then, vice-verse) followed by odd/even middle seats, followed by odd/even isle seats.  Not sure if I will ever be boarded this way, especially because airlines like to let their premier customers board “at any time” and that disrupts just about any organized method.  So, it may be that just letting people on in any order is the right thing to do.

In financial services, this modeling approach is not yet in vogue.  To my knowledge, no vendor is doing this although I wouldn’t be surprised if some of the big banks are using it.  John Geanakoplos, a Yale economist, has developed some prototype models of mortgage prepayments and defaults related to how underwater a homeowner is.  His analysis shows that unless banks write down the principal owed by many homeowners by at least about 20%, those homeowners will eventually default on their mortgages and the banks will ultimately lose a lot more money when they have to pay for the foreclosure process.  Since banks aren’t likely to listen to such advice, he’s trying to get the Federal Government to give banks incentives to do so.  Unfortunately, they aren’t listening either.  John’s prediction?  Some US cities will have a secondary housing crash in the not to distant future because a second wave of defaults is on the horizon.

The more I learn about agent based models, the more I’m surprised it hasn’t taken off sooner.

One Response to Agent Based Modeling

  1. Gabe says:

    Awareness may be the key forward. Without awareness in the research, engineering, management and policy-making fields it might take an unduely longer period of time for the importance of ABMs to come to to light in full beam

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s