Evolutionary Computation Applied to Agent-Based Simulation Modeling of Climate and Social Dynamics

Detailed population distribution data are often unavailable for building spatial agent-based models of climate change effects on humans, especially for past, historical data. A key challenge is to generate approximations of historical (or even future) populations, to initialize our models, based on certain spatial qualities of the landscape. We use evolutionary algorithms (EAs) to tune a relatively simple ”placement algorithm” or ”settlement algorithm.” The EAs generate modern settlements that match LandScan data. When generating historical populations, we include as much information as is available, and then let our algorithm generate missing parts of the spatio-temporal distribution. We illustrate this procedure based on evolutionary computation with the NorthLands model of climate change and social dynamics, using MASON and ECJ. Results are positive and encouraging, highlighting additional research directions. A critical task in developing viable, empirically-based agent-based models (ABM) of complex systems with coupled human, artificial, and natural (CHAN) component subsystems is the specification of initial population distributions. Theory, observation, and experience can provide significant help, but in most cases finding proper population distribution values is a hard, non-trivial task in ABM research. This is especially so for ABMs on climate change and societal dynamics, where spatial and temporal scales are relatively large and theory and observations on population distributions are incomplete. This paper presents a novel methodological procedure for obtaining population distributions in ABMs of CHAN systems using genetic algorithms (GA) from evolutionary computation (EC). The GA-based procedure is illustrated with the recent MASON NorthLands ABM, created to analyze climate change scenarios and societal consequences (Cioffi et al. 2015). Results show the advantage of our GA-based procedure over other approaches, such as manual tuning. The next section provides and introduction with motivation and background on earlier related research, followed by sections describing, experimenting with, 2 J. Bassett & C. Cioffi and discussing our proposed GA-based procedure. The final section provides a summary.