Drivers of Late Pleistocene human survival and dispersal: an agent-based modeling and machine learning approach

Abstract

Understanding Late Pleistocene human dispersals from Africa requires understanding a multifaceted problem with factors varying in space and time, such as climate, ecology, human behavior, and population dynamics. To understand how these factors interact to affect human survival and dispersal, we have developed a realistic agent-based model that includes geographic features, climate change, and time-varying vegetation and food resources. To enhance computational efficiency, we further apply machine learning algorithms. Our approach is new in that it is designed to systematically evaluate a large-scale agent-based model, and identify its key parameters and sensitivities. Results show that parameter interactions are the major source in generating variability in human dispersal and survival/extinction scenarios. In realistic scenarios with geographical features and time-evolving climatic conditions, random fluctuations become a major source of variability in arrival times and success. Furthermore, parameter settings as different as 92% of maximum possible difference, and occupying more than 30% of parameter space can result in similar dispersal scenarios. This suggests that historical contingency (similar causes – different effects) and equifinality (different causes – similar effects) are primary constituents of human dispersal scenarios. While paleoanthropology, archaeology and paleogenetics now provide insights into patterns of human dispersals at an unprecedented level of detail, elucidating the causes underlying these patterns remains a major challenge.