We OR folks believe we support making better decisions. Do we?
If the thesis were correct that canonical OR (which is deterministic optimization) would lead to better decisions in practice, then more than the largest companies would use the technology.
They are not. Even the S&P500 companies only use OR for select applications, usually for transportation and logistics operations as well as revenue management. Our analysis is that traditional OR does, in fact, frequently lead to profit lost instead of profit gained.
Take the airline industry. As one of the early adopters of OR, and an avid investor in the technology, this industry is hardly thriving. Why is it that one-aircraft airlines often find a strong foothold in the industry (see the beginnings of any younger airline) while large airlines, with tons of synergetic potential that OR is supposed to leverage, are frequently struggling and always seem just one winter storm away from Chapter 11?
The inconvenient truth is that deterministically optimized operational plans look awesome on paper. They are even provably optimal. And yet, operational performance is lacking. Why? Obviously, because the future for which we optimized our plans was not the future that then happened. The plans we made were too brittle. They did not hold up in reality.
Alas, airlines lose their shirt not on paper but on the day of operation.
Why are MIP-optimized plans so brittle? Take this post. It describes the canonical view of machine-learning-to-operations-research (ML2OR) pipelines:
- Make awesome forecasts, cross-validate them, and make sure they are as good as they can be.
- Shove those forecasts into a deterministic optimization model, most likely a mixed-integer programming model.
- Implement the optimized decisions minutely.
- Observe a massive discrepancy between the optimized KPIs and the real KPIs.
Where is the gap, you ask? It is quite simple, actually. In the above 4-step failure recipe, note how all uncertainty magically vanishes between Step 1 and Step 2. However, uncertainty cannot be solely dealt with in the predictive step. They must be considered in the optimization itself to obtain operational plans that hold up under real-life deviations from arithmetic mean expectations.
This is why we built InsideOpt Seeker. It is the first general-purpose stochastic optimization solver that allows you to optimize your plans efficiently against tens of thousands of scenarios. Why that many? Because a daily event that can occur with 0.3% probability will, on average, happen once a year. And you want operational plans that are robust. You want your risks mitigated. And you want to understand how risk minimization trades against expected KPI performance.
Your Success Recipe
- Forecast posterior distributions instead of making point forecasts.
- Hand the forecasting model to InsideOpt Seeker. Note that the forecasts may even depend on the decisions to be optimized.
- Find your ideal trade-off between risk and expected reward.
- Observe your operations execute with no major disruptions and high efficiency.
Your first step:
Go to insideopt.com and pip install insideopt-seeker.