Prescriptive analytics, and specifically optimization, has traditionally been treated as a stand-alone domain. This has meant that the inputs for optimization have been manually created, and that the outputs have been produced in isolation from all other systems. In practical terms this equates to increased cost, delays, errors and a frustrating lack of flexibility. For these reasons it is important that the prescriptive analytics environment is integrated into the overall systems environment as fully as possible.
One of the most significant overheads associated with performing prescriptive analytics is the creation and maintenance of business rules, or in the terminology of optimization, the constraints. Even modest optimization projects might involve hundreds or thousands of rules, and recreating them for every optimization problem is a heavy overhead prone to errors. Ideally the prescriptive analytics tools should have access to a business rules data base and management system and be able to convert them into a format that makes sense. In fact for large optimization projects such a facility is not really an option.
Business rules typically express how resources can be combined. It obviously does not make sense to offer a customer a five per cent discount after offering ten per cent on the same product. Neither is it likely to be acceptable that the whole of a workforce is laid off periodically. These and tens of thousands of other constraints typically influence the way large businesses operate, and since prescriptive analytics makes extensive use of them there should be a high level of integration.
However it is not just business rules which should be integrated. Many of the inputs involve the specification of forecasts and other types of prediction. Predictive analytics will generate many models with predictive power – the propensity of customers to respond to an offer for example. These can be used as input to a optimization project, and once again it is extremely useful if the predictive and prescriptive analytics environments are integrated. Optimization caters for probabilistic inputs (a sales forecast for example) through stochastic optimization techniques, and since these facilitate much greater sophistication, the integration with probabilistic predictive models is very important.
Business intelligence, or descriptive analytics, can also provide key inputs for prescriptive analytics, providing information on what has happened and is happening. Information such as the sales of items by region and period might constitute key information in production planning optimization, along with many other metrics. Again it is important that output from reporting and analytics are available to optimization projects.
Finally the results of optimization need to be deployed, and done so in an environment where performance can be monitored, measured and managed. The results of optimization may have a very short life, and it is essential that changes in circumstances can be compared with assumptions made when an optimized model was built and new models built in a timely manner.
A lack of integration is not so problematic for small, one-off projects, but it becomes a major headache as projects grow in size and frequency. Integration should be in the top three or four requirements when selecting a prescriptive analytics platform.
Next article in this series: Prescriptive Analytics Applications