Accuracy and Stability Trade-off in RSSI Window Aggregation for MLP-Based Wi-Fi Indoor Positioning

Wi-Fi RSSI fingerprinting is practical for indoor positioning, but short-term signal fluctuations can reduce classification consistency. This study evaluates how fixed RSSI window size affects the accuracy, stability, and sample availability of an MLP-based indoor positioning system. RSSI from three Access Points was collected at 45 ordered corridor positions. Non-overlapping windows with n=1,3,5,7,9,11,15,21 and 29 were represented by the mean and standard deviation of three RSSI channels. The MLP was assessed using stratified 5-fold cross-validation with supplementary order-blocked validation. Relative to the n=1 baseline Accuracy of 68.24%, n=7 achieved 90.60 ± 0.71% Accuracy, 95.08 ± 1.31% Top-2 Accuracy, 0.9047 ± 0.0060 Macro-F1, and 0.4319 ± 0.1364 Ordinal Error. Larger windows achieved higher mean Accuracy but sharply reduced window-level samples and increased fold-to-fold variability. Order-blocked validation produced more conservative estimates and shifted the strongest performance region toward n=9 to n=11. These findings show that fixed-window selection should consider predictive performance, evaluation stability, sample availability, and validation design rather than maximum Accuracy alone.