Casey Helgeson (Penn State), Nancy Tuana (Penn State)
Other things being equal, simpler theories (or hypotheses, or models) are better. Theses of this form have long been discussed both in science and philosophy (see, e.g., Baker, 2016; Sober 2015). There are many variations on the idea, depending on — among other things — what you mean by “simple” and by “better.” We address the use of a large class of scientific models to inform risk management decisions, and within this context we articulate and defend a new variant of the “simple models are better” thesis. We also discuss associated trade-offs, though how best to balance those remains an open question.
The measure of simplicity that we investigate is the computational cost of running a simulation model on a computer. Simple models require less computation; complex models require more time and/or computing power. We have in mind the kind of computational models — widely used in, e.g., atmospheric and earth system modeling — for which analytical solutions are unavailable and model outputs can be computed only by sequentially calculating subsequent states of the system to describe system behavior over time.
The epistemically important benefit of computationally simpler models is that it is more feasible to estimate uncertainties in the projections of system behavior that they produce. Such uncertainty estimates come from repeated implementations of the model with different inputs. The larger the ensemble of model runs, the better the characterization of uncertainty. So on a fixed computing budget, (computationally) simpler models lead to more thorough exploration and characterization of uncertainties.
In decision-making contexts such as coastal flood risk management, often it is extreme events that are the most concerning, and small differences in the estimated probability of extreme events can sway a decision about how to best manage the risk. Small ensembles of model runs -- a corollary of using computationally complex models -- are particularly badly suited for constraining the probability of such extremes. The result is a coupled ethical-epistemic problem (Tuana, 2017) in which capacity to supply the most decision-relevant information (probability estimates for the most concerning events) trades off with other desiderata that tend to make models more computationally complex (such as increased spatial or temporal resolution, or inclusion of more inputs and processes).
