Meinolf Sellmann

First we run a prescriptive solver to obtain an operational plan, then we put that plan into a simulator to find out what may actually happen? Not only does this imply building two models for the same problem. But what do you do when your risk officers find your prescribed plan lacking? How do you build a back channel to your perscriptive model? That is why you should consider the variability in circumstances that your plan must work in from the start. A solver like InsideOpt Seeker supports this. You can then get the distributions of your KPIs directly from the solver and visualize these. If one or more does not pass the risk analysis, directly modify the model and obtain a solution that fits your risk profile.

This article has ZERO relevance to OR. If such general topics are considered in a specific domain, then the questions asked need to be much more specific than "Are deadlines being missed due to poor planning? Is the team unclear about their roles and deliverables?"

An important aspect that should be added is how to gain trust. Knowing the respective clients’ industry and speaking the lingo are table stakes. A typical trap that OR companies fall into is believing that their predict-then-optimize plans will realize their paper value. In reality, things rarely work as planned. https://www.youtube.com/watch?v=Xr9gPGo-T2g Therefore, it is absolutely crucial to track how the plans prescribed fare in the actual messy operations—and to learn fast. To this end, using the right solver is crucial. Modern solvers like InsideOpt Seeker actively support this stage.

A good path to success is indeed to start with a flagship customer who will provide you with knowledge and typical jargon spoken in their respective industry. That said, it is easy to get locked in this way and to stay a consulting company forever. Even though optimization technology is general-purpose, many companies stay industry-specific for this reason.

Since this article is supposed to be about OR, I miss the decisioning pipeline here. The common practice is to feed estimates directly into optimization models, which is a terrible practice. Why? Because then the optimization model will squeeze every last bit out of that one predicted future, which makes the plan created extremely brittle regarding any deviations that will occur in practice. Instead, set up your ML models to provide (joint) posterior distributions of the data. Then use stochastic optimization in the downstream optimization model to take the uncertainty in the forecasts into account. Modern solvers like InsideOpt Seeker actively support this workflow.

Two truths: 1. Most people positively suck at giving feedback. 2. High performers *do* identify themselves with their work, and *of course* they take it personally! Therefore, if you want to learn how to take feedback, learn how to *give* feedback first. Then, when receiving feedback, you can take the nuggets and turn them into feedback that you can accept and leverage for yourself. To give good feedback: 1. Know what you are talking about. Get the full picture and also gather information from the person you give feedback to. 2. Explain objectively what consequences the action you are providing feedback on had. 3. Coach alternative actions that could have been taken instead. 4. Create the will to do better in the future.