ACTUATE: Adaptive, Multi-Factor balanced, Regulatory Compliant Routing ADM Systems

Fairness is a concept that is central to making equitable decisions about the distribution of critical resources. Traditionally fairness has been studied in resource allocation problems where an optimization function is used to allocate limited resources among few users. In this situation, the goal is to distribute the resources “fairly” among the users. More recently, with the growth in machine learning use cases, fairness has become an emerging topic for AI techniques. Fair machine learning aims the learn decision-making models that are not biased toward any class. The challenge here is to design accurate machine learning methods while providing equitable treatments across all classes or subgroups. While these two problems of prediction and optimization have been studied separately, their combination of a machine learning method to predict key aspects of the formulation of an optimization problem (called predict-plus-optimize) is a new challenge in artificial intelligence. In this project, we study the problem of fairness in predict-plus-optimize problems. In particular, we are investigating the impact of fair learning on optimization problems. For example, in search and rescue planning in response to a natural disaster, there is a need to jointly predict the location and condition of victims of a disaster while simultaneously optimizing the allocation of routing of rescue teams to victims in a fair manner to maximize survival outcomes. We are also examining explainability methods that can help interpret the solutions of predict-plus-optimize problems.