Addressing Problem Drift in UNHCR Fund Allocation

Optimisation models are concise mathematical representations of real-world problems, usually developed by modelling experts in consultation with domain experts. Typically, domain experts are only indirectly involved in the problem modelling process, providing information and feedback, and thus perceive the deployed model as a black box. Unfortunately, real-world problems “drift“over time, where changes in the input data parameters and/or requirements cause the developed model to fail. This requires modelling experts to revisit and update deployed models. This paper identifies the issue of problem drift in optimisation problems using as case study a model we developed for the United Nations High Commissioner for Refugees (UNHCR) to help them allocate funds to different crises. We describe the initial model and the challenges due to problem drift that occurred over the following years. We then use this case study to explore techniques for mitigating problem drift by including domain experts in the modelling process via techniques such as domain specific languages.