MODELING HUMAN MIGRATION AND POPULATION GROWTH WITH DEEP LEARNING AND MESOSCOPIC AGENT-BASED MODELS

Abstract

Modeling human migration and population dynamics is vital for governments and social scientists so that they may effectively prepare jobs and living spaces for influxes of people fleeing war, famine, climate change, or discrimination, along with those simply seeking a better economic standing. Previous work on the topic centers around gravity and radiation models for immediate migration prediction; this thesis proposes a methodology for forecasting long-range time series of population and migration data using country specific successor-state neural network models that act alongside a separate migration and distribution model. Further, two types of successor-state models are considered: a feature-feature model that explicitly uses observable features, and a feature-statement model based off of gated recurrent units that makes predictions using both a feature set and a hidden state. Using the World Development Indicators and Global Bilateral Migration datasets from the World Bank, the models are able to successfully forecast populations with reasonable death and birth rate predictions on most countries; however, the two-model system proves unable to reliably predict international migration flows. Future work on the proposed modeling systems should aim to unify the successor-state and migration distribution models to rectify the migration discrepancies predicted by the two-model system.