Tipping points are an active and growing interest in both the scientific and political study of climate change: what are they, how can we identify them, and how can we avoid them (negative tipping points) or encourage them (positive tipping points). As climate change worsens, scientists and policy analysts have turned to computer models of complex, interconnected, human-Earth systems to help understand and address both its physical and social aspects. As this field has matured, so too has the complexity of both the models being developed and the questions being asked with them. Agent-based modeling (ABM) is one framework that has become popular for its ability to observe system-level behaviour without closed-form equations due to its encoding of heterogeneous individual-level behaviour.

Analyzing the data generated by ABMs is not straightforward, as they tend to have many input and output dimensions, most outputs are either temporally or spatially distributed (or both) and can be sensitive to stochastic effects. When applying a exploratory modeling or deep uncertainty lens—a philosophy that seeks to explore the effects of assumptions made in a model’s development and parametrization, understanding more about the modeled system’s behaviour as opposed to attempting to predict it—the complexity of this analysis grows further. However, this complexity should not discourage analysts from bringing existing Decision-Making under Deep Uncertainty (DMDU) methods to ABMs.

This study applies one such method (scenario discovery) to a complex ABM of household and firm climate adaptation in a coastal economy, attempting to uncover the existence of socio-environmental tipping points in the system. Based on a previously developed analogy connecting the output space of an ABM to the traditional notion of a physical phase diagram, Scenario Discovery is used to generate such a phase diagram and infer tipping points at the boundaries between distinct system states. Ultimately, a set of possible population-change tipping points are generated.

While this work demonstrates the fitness of scenario discovery as a tool for exploring the output spaces of ABMs and finding tipping points within them, it is very preliminary. The work should be repeated with several improvements. First, either the uncertain parameters varied in this study should be selected to be more policy-relevant, controllable system factors, or the system states and thus the tipping points should be expressed in terms of endogenous variables instead of input parameters. Second, this study demonstrates that the typical approach to processing stochastic replications in exploratory modeling—simply averaging all outcomes—is not fit for use with complex modeling like ABM. Despite the computational and cognitive load introduced by simultaneously handling both a wide uncertainty space and many stochastic replications, efforts must be made to ensure any dynamically distinct behaviour generated by the original model is not lost to averaging. Studies like this one that do not put in this effort risk enabling the extraction of incorrect political and policy lessons.