This morning I read a research paper titled “Designing
a Distribution Network for a Soda Company: Formulation and Efficient Solution
Procedure” by Arguello et al.

A real-world business problem was leveraged in the research. A soda company
was running into distribution challenges due to a recent reconfiguration of
manufacturing strategy and footprint and a change in the pre-sales scheme on
the commercial side. The paper proposes to use a Lagrangian relaxation
technique (developed in the early 1970s by Held and Karp to solve the traveling
salesperson problem) to address the challenge.

Overall, the proposed solution was an improvement upon the classical
approach. However, since orders in the real-world flow constantly, and
optimization models, by their very nature, use data snapshots at one point, it
is a challenge to use them as tactical models.

Even in this scenario, with the scale of operations, using an optimization
model to plan in real-time, since the delivery agent’s pickup needs to happen
on the day of the order, to be delivered after two days. I emphasize that deep
learning algorithms are critical to helping make optimization models more
tactical. That will allow optimization models to be leveraged for day-to-day
planning in near real-time.

The idea is simple and has been incorporated into products from simulation companies like Simio (Reference). We are creating thousands of scenarios to generate big data sets for training AI models. These models can simply look at orders and suggest assignments and routes without running optimization models. That same approach can be extrapolated in optimization.

The critical point is that the optimization model still churns out
prescriptions for scenarios. And even when a deep learning model is being used in production, the optimization model keeps running in the background to incorporate new scenarios and data. All the output from the continuous optimization run becomes a constant feed of learning data for the deep learning algorithm.

The result is a model that can provide near-real-time recommendations powered by optimization data. This combination can help address the run time lag of optimization models that is often a bottleneck in leveraging them as near real-time tactical decision-making models.