For those that have worked in the area of operations research, mathematics or computer science the traveling salesman is an intriguing and well known problem. The problem sounds very easy, what is the shortest route traveling from one destination to the next and returning to the place you started. What makes this problem intriguing is that no mathematical algorithm exists that guarantees the shortest route in an acceptable amount of time (to be precise: it’s an NP-Complete problem and as such the most efficient algorithm to date – and probably forever – solves the problem in a time that increases exponentially with the number of cities).  In practice this means that normal CPU’s would need days  or even years to solve problems with a relatively small number of cities (e.g. 100).

Applying marketing data includes many optimizations for which we can prove that no efficient algorithm exists. Examples are television reach optimizers or allocating inventory towards brands. For these problems, as a real optimal solution cannot be guaranteed, mathematicians need to develop heuristic algorithms; an algorithm that tries to get to the best possible solution in a limited amount of time.

And that brings me to the key subject of this blog. The deflation I’ve seen on what it means to optimize. Ten years ago, my clients asked me about how our optimization techniques work; they even got independent mathematicians to audit our methods. Nowadays this doesn’t happen anymore. To have a button in software that says “optimization” seems to be enough to convince prospects. And this is a slippery slope as some optimization routines I’ve seen has moved towards simply users clicking (with some basic information) what they think is best. And let me guarantee, users are not able to optimize as good as good old-fashioned mathematical optimization techniques.

Don’t believe me, please download a little piece of software I wrote many, many years ago. This software allows you to try and find the best route for a traveling salesman before the computer gives it a go. With about 40 cities, on average – in my case – the computer does 8 to 10% better. That’s a lot if it is a 10 mln dollar decision!

And yes, as the computer algorithm cannot guarantee the best solution, you might be able to beat the algorithm. Do let me know (with screenshot!) if you’ve been able to do this.

People sometimes ask me what I think about “agent based modeling”. As a mathematician it’s actually somewhat confusing that during my studies I learned about Markov processes, simulating queuing processes, Bayesian statistics. But it seems that the good old world of mathematics has caught up with the world of branding: genetic algorithms, neural networks, agent based modeling, etc.

So let’s put agent based modeling in the box it belongs. It’s nothing more (or less) than simulation. I’ve studies queuing processes by applying simulation models. People would come into the process, each would determine which queue to select and the time it takes to serve a person was based on some random distribution. Without me knowing it, I was applying agent based modeling.

For marketing purposes, simulation I do believe has great potential. We can research individual behavior and decision rules but a system to aggregate these individual characteristics and to allow for scenario planning based on such an aggregation would be far too complex to execute without a simulation model. But with a simulation we could have ‘agents’ seeing ads, using the internet, going shopping, etc. If current researched behavior leads to current brand outcomes, it seems like a credible step to change the environment of these agents and to see how they would react. It’s like creating a virtual environment where you can give virtual consumers stimuli and to see what and when they’ll purchase.

So what’s next? Consumers learn and change their behavior and decision rules based on historical stimuli. This means that in our simulation model we would need to allow our agents to learn. So the next big thing is “agent based artificial intelligence”! Just googled this and very disappointed to see this already gets 62,000 hits. At least I do believe I can claim to be the first to bring this idea to the marketing community.

By the way, even though I’ve studied queuing process, I too always elect the wrong queue!