Evaluating rule induction algorithms for scenario discovery

Scenario discovery translates large simulation ensembles into interpretable input regions linked to policy-relevant outcomes. However, comparisons of SD algorithms remain ad hoc and hard to reproduce. We propose a general workflow to evaluate rule induction methods for SD. The workflow (i) provides synthetic benchmarks that expose axis and directional misalignment, nonlinearity, boundary fuzziness, and dimensional noise; (ii) unifies metrics and diagnostics around coverage–density trade-offs, interpretability, runtime, and scaling; and (iii) prescribes a staged experiment design from low-dimensional screening to stress testing. We illustrate the approach by comparing established algorithms PRIM and CART with an oblique decision tree variant called HHCART(D), finding that the latter does not outperform the former. Our workflow surfaces method-specific trade-offs and supports principled, reproducible algorithm selection for scenario discovery. • Reproducible workflow for evaluating rule induction methods for scenario discovery. • Synthetic test shapes stress axis rotation, nonlinearity, and noise. • Unified metrics: coverage, density, interpretability, runtime. • Open code and benchmarks enable fair, extensible comparisons.