New Analytic Approaches in Decision Optimization

There’s a lot you can do right now to optimize your decisions. You can model your decision-making to understand it better, experiment to gather data about what works and for whom. You can engage in continuous improvement, making small changes regularly. And you can optimize your decisions mathematically.

But what will you be able to do soon? What does the future hold for decision optimization?

The field of machine learning and artificial intelligence (ML and AI) is exploding with new developments. There’s obviously tremendous excitement about ML along with some valid concerns about data bias, overfitting and “black box” models. Overall, though, the potential for ML and AI to develop more accurate and less uncertain predictions more quickly is clear. These models will improve any optimization that uses them.

Of particular interest are developments focused on Causal Inference. Causal Inference or Action Effect models answer the question “what if we did x or y or z? “. They predict the likely outcome of different decisions, enormously helpful when trying to pick the best of several different potential decisions. They don’t eliminate the need for experiments, but they do add a great deal of value when optimizing decisions.

Decision trees are often used in decision management and decision optimization. The standard algorithms for developing trees are well established and effective. But they tend to assess each variable to find good splits one node or level at a time. This is a “greedy” approach and can result in overfitting or local optima because the overall predictive power of the tree is not being considered – just the next node.

Research in this area by MIT, Microsoft, IBM and FICO is focused on creating the entire decision tree at once and considering all the potential combinations of variables and splits to find the best. Improvements in computing power meanwhile are allowing the many combinations to be efficiently be considered. This should result in smaller, easier to manage, easier to explain decision trees that are more predictive.

In the end though, we need to move from machine learning to business learning. Today we rely on specialists to do advanced decision modeling, to build analytic models and to configure optimization engines. Business users are already reviewing experiment outcomes, modeling decision requirements and managing business rules. They’re not going to stop there. They’ll want to be able to manage and control the whole decision optimization environment.

Easy to use tools increasingly let them manage their data feeds and build machine learning models especially in the context of specific well-defined scenarios. Soon this will evolve to allow domain experts to define their objectives in business terms. The optimization engine will automatically run various scenarios for the domain expert to choose between. Running these scenarios as experiments will create new data and automatically update the next iteration. A seamless combination of automated machine learning, decisioning and optimization technologies. Automated business learning.

But don’t wait, get started optimizing your decisions now using the approach outlined in our new paper. The more you know, the further you are along, the easier these exciting developments will be to adopt and the further ahead you will get.

Other posts in this series: