My first foray, over a decade ago, into agent based modeling (ABM) was developing one as a member of store operations for a specialty retailer in Columbus, Ohio. The objective was to model shopper and associate behavior in a store with the primary intent to derive the optimal shopper-to-associate ratio by brand, time of year, and day of week.
Consulting the latest edition of the “bible of simulation,” Simulation Modeling and Analysis by Averill Law, I was pleased to find that after interviewing many experts, Dr. Law’s conclusion was that ABM “is just a special case of DES” (DES being discrete event simulation). If that held true, then building shopper models for retail simulation should be possible in a general purpose DES package. When I started, I consulted an institute of higher learning in Dayton, Ohio. Not only was I told I couldn’t do it but that it couldn’t be done in Arena…I could and did. Don’t you just love experts?
Reneging, or jockeying, has long been an entity behavior that has been modeled since DES packages came into existence. Balking has as well. Aren’t reneging and balking indicative of an agent’s behavior? The entity has a propensity to wait (or not), is aware of what is occurring in front of itself in line and beside itself in other lines then makes decisions based on that information. In my opinion, that is autonomous behavior and decisions are being made based on the environment.
For this application of ABM, there are two types of agents each with their own modeling complications. First, there is the shopper. Is the shopper moving with intent to obtain a predetermined list of items or just browsing? How fast does a shopper move in each mode? What if it’s a male versus a female shopper? An adult pushing a cart or stroller? Behavior in the store will also be driven by the shopper’s a priori knowledge of the layout and location of the products in the store. Finally, there’s the wrap desk or checkout area. Some retailers intentionally place the wrap desk in the back of the store. No sense discouraging shopping with a highly visible line that everyone passes when they walk into the store. Contrast this with your typical big-box retailer or grocery store. All shoppers enter those environments and the first thing they observe is whether or not there are lines. Also note how Lowes, Home Depot, and Kroger attempt to conceal the lines with fixtures, products, and partial walls.
The second type of agent in this scenario is the associate. In specialty retail, providing a great customer experience is generally more important than in big-box retail. When was the last time you were able to quickly and easily find an associate in a Target store? The associate will also have different roles that drive different behaviors and decisions. Is the associate in selling, stocking, or checkout mode? If they are in selling mode, what do they do with their down time? When a customer enters the store, does the associate observe or become predatory with rapid fire, probing questions, “How can I help you? What brings you in the store today? Would you like a bra fitting?” Recall how perfume testing turned into perfume tasting in Elf starring Will Ferrell when he was walking thru the Gimbles store not even shopping. How do the associates work in concert? As the shopper moves to another room or department, are they engaged by another associate? Do they frustrate the shopper by asking the same sequence of questions? What happens when a shopper needs help in a fitting room? For some retailers, the last thing the retailer wants is for the shopper to leave the fitting room to return to the floor and look for the correctly sized products themselves. What staffing levels and associate behaviors deliver the customer experience the brand is driving towards? We can test these things in real world situations or we can build a model to analyze and then eliminate the dysfunctional, brand-damaging behaviors.
Data collection for these types of projects, as with any other, can be extremely time consuming. When I built that first model, the retailer happened to have traffic counters. So, in addition to point-of-sale data, I only had arrival times and average shopping durations by day and that was it. I had to observe the shopper and associate behavior while in the store. Today, with systems such as IBM® Presence Insights, shopper and associate behaviors can be tracked extensively and the data needed to drive a model can be derived. These systems allow different zones to be sensored and tracked. The system logs smartphone information when it attempts to connect to the store’s Wi-Fi or if they have a store-specific app on their smartphone. Imagine sensoring the threshold of each bathroom or fitting room. With that information, each event with that device could be tagged as male or female. In addition, more rigorous analyses can be conducted regarding reneging and balking behaviors as shoppers move in, out, and by wrap desk zones.
One of my favorite quotes is “It’s the archer, not the arrow.” In hindsight, I was confident that a general purpose DES package could be used to build an ABM, and I’m glad that I proceeded without a procurement cycle for an ABM specific tool. At the end of the day, the tool is merely an enabler. The modeler makes things happen.