Extrapolation-Aware Gaussian Process Configuration for Small-Sample Complex-Equipment Cost Estimation

Complex-equipment cost estimation often relies on very small historical datasets, and new targets may fall partly outside the feature ranges represented by those data. This study examines whether two observable data-state indicators can provide a transparent basis for configuring Gaussian process regression (GPR) before the target cost is known. The proposed ED-AGPR framework combines a target-specific parameter extrapolation ratio with a case-level sample-to-feature ratio to route each target among three predefined GPR configurations. The empirical assessment retains five source-defined application targets and uses case-specific leave-one-out cross-validation across the three datasets (29 pseudo-target predictions in total), matched component contrasts, conventional and composite GPR baselines, five-seed optimizer-robustness analysis, threshold sensitivity, and predictive-interval coverage diagnostics. Under leave-one-out evaluation, ED-AGPR achieved a mean absolute relative error of 48.51%, compared with approximately 52% for standard Matérn-3/2 and RBF GPR baselines. The paired case-cluster bootstrap descriptive ranges for those standard GPR differences included zero; the component contrasts showed that no single modeling ingredient was uniformly beneficial, and the predictive intervals substantially under-covered their nominal levels. The results therefore support ED-AGPR as an auditable empirical configuration heuristic for the three examined small-sample datasets rather than as an estimator with established predictive superiority or universal deployment readiness.