Machine Learning Models for Fire Incident Response Optimization

Fire departments operate under competing demands: they must arrive fast, reduce property loss, and identify suspicious incidents early, often with limited resources and uncertainty about which operational factors matter most. NFPA 1710 sets a benchmark of 240 seconds (4 minutes) for first-engine arrival, yet real conditions make consistent compliance difficult. Decision-makers lack transparent tools to understand which combinations of incident and operational conditions enable target outcomes, and what trade-offs exist between competing objectives. This paper proposes an audit-friendly decision-support framework that combines machine learning (ML) with Global Multi-Objective Optimization (GMOO). The central innovation is inverse design: rather than traditional forward prediction ("given inputs, predict outcomes"), we ask the inverse question—given desired outcomes (low loss, high arson probability, fast response), which incident and operational input configurations make them achievable? This reframing transforms ML models from passive predictors into active tools for scenario exploration, trade-off quantification, and model auditing. Using the Colorado NFIRS 2024 dataset (777,930 records; 550 features), we train three predictive models, each representing a core operational outcome that fire agencies must manage. The fire severity model estimates expected property loss using a two-stage architecture that handles the heavy mass at zero loss (R²(log) ≈ 0.79 on non-zero loss subset; MAE = $6,288 on full test set, n = 4,118). The arson risk model predicts the probability of intentional ignition, implemented as a CatBoost–XGBoost ensemble with explicit label-leakage prevention (F1 ≈ 0.81, AUC ≈ 0.88). The response time model estimates alarm-to-arrival time, enabling scenario testing against NFPA targets (MAE ≈ 2.44 minutes, R² ≈ 0.72). GMOO treats these trained models as objective functions and systematically searches a metadata-defined raw feature vector (24 features derived from the severity, arson, and response-time model inputs, plus time parents) under typed variables and empirical bounds. Each configuration represents a specific combination of input values that produces predicted outcomes across all three models; configurations satisfying all target constraints are termed "hits." Through progressive experimentation (single-objective → 2-output → 3-output → 4-output), GMOO discovers interpretable configuration sets: 908 high-arson configurations (max probability 0.9994) and 905 low-arson configurations (min 0.0058) demonstrating clear model discrimination; 605 severity-optimized hits with minimum loss approximately $1,026; and 225 multi-objective hits across three response-time targets, with the best achieving a 1.02-minute predicted response time and 17 configurations meeting NFPA 1710 compliance. The framework makes trade-offs explicit for auditing and planning. Pareto analysis reveals a non-linear relationship between response time and loss, with a practical balance region in the 2–4 minute range. Beyond trade-offs, the framework surfaces actionable patterns: trash/rubbish fires comprise 20.4% of high-arson configurations with nighttime and weekend concentration, while NFPA-compliant configurations show higher self-reliance (AID=0) (35% vs. 8%), indicating reduced mutual-aid dependency. Constraint cost quantification shows that adding operational constraints increases minimum achievable loss from $869 (unconstrained) to $2,522–$4,295 (multi-objective), allowing planners to see exactly what each constraint costs in terms of other objectives. Critically, the inverse optimization process also functions as a model audit: when GMOO identifies configurations with implausible predictions—such as sub-two-minute response times that violate physical turnout-time constraints—these findings expose specific deficiencies in the underlying ML models rather than representing operational recommendations. This dual role—operational insight where models predict within their valid range, model diagnostic where they do not—is a distinctive contribution of the framework. GMOO results serve a dual purpose. Where model predictions fall within operationally plausible ranges—such as the 2–4 minute response balance region and arson indicator profiles consistent with known ATF patterns—they constitute actionable decision-support findings. Where predictions are implausible—such as sub-two-minute response times that violate physical turnout constraints, or a $133 arson damage premium far below observed differentials—they function as model audits, exposing specific deficiencies in the ML pipeline that warrant investigation and guide model improvement. Both categories are contributions of the framework. Results are Colorado-specific, and real-world factors including staffing availability, weather, and concurrent incidents remain unmodeled.