At a glance:
- Many businesses use machine learning and artificial intelligence to build forecast models that help inform business-critical decisions.
- However, the world is an unpredictable place, and we shouldn’t make decisions based on our single best guess of the future. Instead, we need to take multiple scenarios into account.
- We look at how stochastic optimisation can be used with forecasting models to make decisions that are more robust to uncertainty, and give organisations the edge when it comes to making business-critical decisions.
As humans, we make thousands of decisions every day. Most of these decisions we can make without much serious thought, using the facts and experience we have assembled over a lifetime to take the correct course of action. More complex decisions, particularly those that require us to anticipate the outcome of future events, can be much more difficult to make.
Today, many businesses use machine learning (ML) and artificial intelligence (AI) to build forecast models that are capable of predicting the likely future. Many of these models use historic data samples to offer one likely view, which can then be used to make a whole host of business-critical decisions. But what happens to these models when the future does not unfold in the way we had expected?
When we enter uncharted waters, where historical data no longer reflects the current circumstances, we cannot expect machine learning models to correctly predict future outcomes. Machine learning can certainly help provide a forecast of the future but, when the stakes are high, it is unwise to make a critical decision based on our single best guess. Forecasts are significantly improved if we account for uncertainties and consider multiple scenarios of what the future might look like.
However, this begs the question; how can the various scenarios be analysed to present organisations with a course of action that achieves the most positive outcomes to business-critical decisions?
This is where optimisation comes in.
Optimisation can be used with forecasting to analyse available data, exploring various ‘what if’ scenarios, and then present the optimum course of action. For example, when trying to forecast the most cost-effective route from A to B while navigating all required stopping locations, optimisation methods can be used to analyse every potential route and delivery sequence to reveal the optimum solution. However, as classic (deterministic) optimisation problems consider one single view of the future, they are not always robust to the uncertainty that businesses face every day.
Combining optimisation and forecasting, to make decisions in an uncertain world, can give businesses a significant edge.
Thanks to the rapid development of optimisation techniques and the improved performance of commercial solvers, there are now some very powerful tools available to businesses facing great uncertainty. Stochastic optimisation, which considers multiple scenarios to produce a decision that is more robust to uncertainty, is one such tool.
What are the practical uses of stochastic optimisation?
To illustrate the breadth of capability of stochastic optimisation, consider some of the questions supermarket chains have been wrestling with this year and how forecasting and optimisation, more robust to uncertainty, could make a difference.
- Where should new stores be located to maximise revenue without detriment to existing stores?
- What jobs should commercial staff be allocated to meet uncertain demand while maximising customer experience?
- How can route scheduling be optimally managed when new requests are being received in real-time?
- How can profits be maximised when cultivating produce on finite land and where both crop yield and future price are uncertain?
- How can the stockpile of commodities (raw materials or utilities, for example) be managed when faced with uncertainty in demand and price?
All of these decisions involve a large degree of uncertainty and a huge number of variables. Combining optimisation and forecasting, to make decisions in an uncertain world, can give businesses a significant edge.
While we have considered decisions facing supermarket chains in the examples above, organisations of all types will of course have business-critical decisions to make in the face of uncertainty. Next, we consider in more depth how stochastic optimisation can help an organisation consider a complex problem and deliver a solution that is more robust to this uncertainty.
Case study: The power of stochastic optimisation
Imagine you manage a large commercial office building and after some analysis of electricity costs, you have decided that you are spending too much. How can you reduce costs and save money?
By attaching a battery to your building, you can decide whether energy demand over time is met from battery storage or from the grid. To reduce costs, the optimal way to power the building is to charge the battery when the electricity price is low (powering the building from the grid during this period), and to discharge the battery to power the building when the electricity price is high. You could even offer to sell a small amount of the battery’s capacity to transmission system operators (TSOs), who use highly responsive battery reserves to keep the overall electricity network stable. This would provide a good source of income but would require decisions to be made over how much battery capacity to make available, and when.
Deciding when to charge and discharge the battery, and when to offer capacity to TSOs, has an inherent level of future uncertainty, including both the future cost of electricity and demand from the building.
To help make a decision, you could use a single forecast (typically an average forecast) to predict what the electricity prices and demand will be over the next 24 hours. However, this method of deterministic optimisation risks making a decision that will be potentially harmful if price and demand are not as forecast. For example, a forecast may predict midday as the highest price for electricity and so suggest discharging the battery during this period to avoid the high price. If two hours later the prices are higher still, however, you now have both an empty battery and higher electricity costs to pay.
Using stochastic optimisation, it is possible to forecast multiple price, demand, and TSO capacity usage scenarios, and to then make a decision with considerations to uncertainty. As well as comparing deterministic and stochastic approaches, it is possible to benchmark performance against a situation without certainty. In this unrealistic ‘perfect’ situation, the system knows all the price, demand, and TSO capacity values ahead of time and reacts accordingly.
Watch the partnership webinar we hosted with Gurobi: Making critical business decisions in the face of uncertainty.