Explanation-Guided Rule Revision for Adaptive Symbolic Controllers

Introduction: Adaptive controllers in agent-based simulations can be evaluated through trajectory diagnostics, but those diagnostics rarely support transparent revision of controller logic. This study tests an explanation-guided rule-revision layer that converts a diagnosed controller failure into an auditable symbolic edit.

Methods: Policy actions were represented as defeasible rules with explicit predicates, conflicts, and priorities in a stylized emissions-regulation agent-based model. The baseline was an uncoached symbolic controller lacking a relaxation rule. Forty fixed diagnosis seeds (1000–1039) were separated from 40 non-overlapping held-out evaluation seeds (2000–2039), which were not used for threshold specification, template selection, or rule revision. Policy pressure was a bounded regulatory-control signal. A predefined template added a relaxation rule when over-conservatism occurred without cap violations. Tracking error measured absolute distance from the emissions cap, whereas conservatism gap measured excessive distance below the safe operating region. Post-burn performance was evaluated using these metrics together with violation rate, overshoot, policy volatility, and failure recurrence.

Results: On paired held-out runs, conservatism gap decreased from 6.952 to 0.770 and tracking error from 10.667 to 4.241. Mean paired improvements were 6.182 (95% CI 5.073–7.303) and 6.426 (5.250–7.608), respectively. Over-conservatism recurrence fell from 0.875 to 0.000, an improvement of 0.875 (0.750–0.975). 

Confidence intervals used 5,000 nonparametric percentile bootstrap resamples of the 40 paired seed-level differences. Violation rate, mean overshoot, and policy volatility remained zero. Local robustness checks across three conservatism thresholds and four initial policy signals preserved the principal result.

Conclusions: A finite, template-constrained coaching layer can connect diagnosis, explanation, symbolic rule revision, and held-out re-evaluation thus supporting auditable local repair of a diagnosed symbolic-controller failure without introducing secondary control failures. Evidence is limited to one stylized environment, one diagnosed failure class, and one predefined revision template; it does not establish global optimality or unrestricted autonomous repair.