Machine Learning for Decision Making under Deep Uncertainty

By Adolfo De Unánue

During the last couple of decades we have witnessed an increase of Data Science applications to almost all kinds of practical problems, or at least that’s what we have been told. However, if we look closer to this undeniable accomplishment, we will discover that the most successful applications are concentrated in what could be classified as “operational” or “tactic” decision making: “This credit card transaction, is a fraud or not?”, “This particular buyer, will buy again?”, “This viewer will like this movie or not?”. As you can see, they have the following characteristics: they are implemented (a.k.a. “running in production”, i.e. the decision triggers an act that modifies the reality), they also need a set of labeled data set, but equally important, they classify[1] an individual in a reduced set of states, called labels. This means that a successful implementation of machine learning, requires (at least) data and a reformulation of the problem in a yes/no question[2].

Although implementations were mainly concentrated in the private sector, these techniques were eventually adopted to the public sector, but their adoption was not identical across the board. We can classify the implementation of machine learning to public policy problems into two groups, depending on the particular problem-solving vision: (1) those who try to solve old econometric problems with machine learning (e.g. replacing a logistic regression with something “newer”, like random forests); and (2) those who applied machine learning, adapting the successes and lessons learned from private sector, to tackle public operational problems with the goal to implement actions with direct social impact at the individual level.

Public policy problems have the following attributes: The stakes are higher, the problems are less defined and constantly evolving, and there are multiple actors and stakeholders without a clear or shared definition of what is the correct thing to do, which techniques or models are valid to use, or how to measure the success of the actions. Public policy entails solving strategic problems with high levels of uncertainty. Our most pressing social problems (infrastructure, climate change, societal organization, migration, crime prevention, etc.) have those characteristics.

Data science and machine learning cannot provide general solutions to these big-picture problems. The state of art is simply not there. Nonetheless, we a have a bunch powerful techniques, lots of data and cheap and accessible powerful computers that we can apply to operational/tactical problems. One of these techniques is DMDU.

DMDU techniques use many models at the same time, not for prediction purposes, but to test the viability/resilience/robustness of different strategies in the face of deep uncertainties. There are at least three cases in which DMDU benefits from Machine Learning:

  1. data gathering (for sensing the environment, which is key in monitoring the plan in the technique Dynamic Adaptative Policy Pathways),
  2. understanding the models’ output (used in Scenario Discovery, for example), and
  3. expanding DMDU models (which are traditionally physical and econometric) to machine Learning models.

In the forthcoming DMDU annual meeting, we have a fine selection of works related to the use of Machine Learning techniques in DMDU.

First, Irene van Droffelaar and Jan Kwakkel, both from TU Delft, will present a novel technique for Scenario Discovery based in Markov Chain Monte Carlo, called DREAM. This technique efficiently samples uncertainty spaces with high dimensionality. They will show us some examples to compare the performace of DREAM, Latin Hypercube and PRIM.

Next, Fernanda Sobrino, from the University of Chicago, will talk about her use of Machine Learning techniques, in a particular novel natural language processing application, using Deep Learning for creating data sets that can be used for further analysis.

Last but not least, Michelle Miro and James Syme, from RAND (cradle of one of the most powerful DMDU methods: Robust Decision Making or RDM), will show us the application of Machine Learning to a very important study on water resources planning in southern California.

All of our speakers will be presenting different points of view regarding the application of Machine Learning in DMDU. Please join us for this exciting discussion on November 12!

[1]I acknowledge that there are other types of machine learning techniques like regression, or unsupervised learning. But those have a more modest success compared with the classification applications.

[2]I am ignoring, for the sake of brevity, the implementation part, the monitoring of the model’s behavior, model selection, model governance, etc.

About this Blog Post:

This blog post is part of a series of posts contributed by the chairs of the 2020 DMDU Annual Meeting. For more information about the Annual Meeting, including registration, visit our website at We hope to (virtually) see you soon!

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.