Flee 3: Flexible agent-based simulation for forced migration

Forced migration is a major humanitarian challenge today, with over 100 million people forcibly displaced due to conflicts, violence and other adverse events. The accurate forecasting of migration patterns helps humanitarian organisations to plan an effective humanitarian response in times of crisis, or to estimate the impact of possible conflict and/or intervention scenarios. While existing models are capable of providing such forecasts, they are strongly geared towards forecasting headline arrival numbers and lack the flexibility to explore migration patterns for specific groups, such as children or persons of a specific ethnicity or religion. Within this paper we present Flee 3, an agent-based simulation tool that aims to deliver migration forecasts in a more detailed, flexible and reconfigurable manner. The tool introduces adaptable rules for agent movement and creation, along with a more refined model that flexibly supports factors like food security, ethnicity, religion, gender and/or age. These improvements help broaden the applicability of the code, enabling us to begin building models for internal displacement and non-conflict-driven migration. We validate Flee 3 by applying it to ten historical conflicts in Asia and Africa and comparing our results with UNHCR refugee data. Our validation results show that the code achieves a validation error (averaged relative difference) of less than 0.6 in all cases, i.e. correctly forecasting over 70% of refugee arrivals, which is superior to its predecessor in all but one case. In addition, by exploiting the parallelised simulation code, we are able to simulate migration from a large scale conflict (Ukraine 2022) in less than an hour and with 80% parallel efficiency using 512 cores per run. To showcase the relevance of Flee to practitioners, we present two use cases: one involving an international migration research project and one involving an international NGO. Flee 3 is available at https://github.com/djgroen/flee/releases/tag/v3.1 and documented on https://flee.readthedocs.io.