A systematic evaluation of machine learning algorithms for WiFi RSSI fingerprint-based indoor positioning under consistent Bayesian optimization

Abstract

Indoor positioning has become essential for location-based services in smart buildings, yet Global Navigation Satellite Systems cannot provide reliable positioning indoors due to signal attenuation and multipath interference. Existing machine learning approaches for WiFi fingerprinting suffer from fragmented evaluation: studies compare limited algorithm subsets using inconsistent hyperparameter optimization and varied metrics, preventing reliable conclusions about algorithm selection. This paper presents a systematic comparison of six machine learning algorithms spanning four distinct paradigms—K-Nearest Neighbors (KNN), Weighted KNN (WKNN), Support Vector Machine (SVM), Random Forest (RF), XGBoost, and Deep Neural Network (DNN)—for WiFi Received Signal Strength Indicator fingerprint-based indoor positioning under consistent Bayesian optimization. All algorithms undergo identical Gaussian process optimization with 50-iteration budgets, ensuring fair comparison under consistent conditions. Evaluation encompasses a 2,632 m² real-world environment with 517 reference points and 12 access points. Results demonstrate that Random Forest achieves optimal performance with 0.7170 m Root Mean Square Error (RMSE) and 94.73% classification accuracy, representing 21.7% improvement over baseline KNN (0.9152 m) and 42.2% improvement over DNN (1.2404 m). Statistical significance testing using Wilcoxon signed-rank test with Bonferroni correction confirms all pairwise differences (p < 0.001) except WKNN-KNN. These findings demonstrate that traditional machine learning can outperform deep learning for moderate-scale WiFi fingerprinting, providing evidence-based guidance for indoor positioning system deployment.