Online

Andrew, Kevin Allen

Computer Science

2023

M.S.

Agent-based models are becoming increasingly useful in studying the behavior of real-world complex multi-agent systems; however, one of the outstanding challenges in the modeling of coupled natural and human systems is the dearth of techniques for creating agents that are able to learn from their past failures and successes, as well as compounded environmental and social uncertainties. This research has been focused on the integration of traditional agent-based modeling with machine learning methodologies for modeling agent decision-making and its recursive impacts on economic, environmental, and societal outcomes, feeding into the dynamic co-evolution of the coupled natural and human system state variables within simulated worlds, resulting in the development of two models incorporating and exploring the use of deep reinforcement machine learning as a driver for decision-policy making in agent-based models. The first of these models is a model of agricultural land use and the adoptionof agricultural best-management practices by farmers in response to ecological and economic scenarios as a result of municipal regulation and variance in the occurrence of extreme weather events. The primary study area used for the model is a region of the Missiquoi Bay Area of Lake Champlain in Vermont, containing 480 farmer agents corresponding to agricultural land parcels within the region. A parameter sweep and sensitivity analysis on model hyperparameters was conducted to explore the effects of changes to agent calibration and training on agent decision-making and model performance. The second model expands upon the scope of the first, including foresteragents and commercial and residential urban agents within a larger region of the Lake Champlain Basin of Vermont. Additionally, the impacts of agent decision-making take place on the simulated landscape, resulting in gradual land cover change over time. Land cover data from the United States Geological Survey's National Land Cover Database was used for initial parameterization, calibration, and training of the model (years 2001, 2006) and model testing (year 2011). Results suggest that with appropriate scoping and hyperparameter selection,the integration of deep reinforcement machine learning techniques into the development of agent-based models can increase predictive accuracy in the modeling of real-world phenomena; however, these gains must be weighed against the increased technical complexity of such a model and the associated risk of introducing model error.