TRA (Track/Rail Algorithm): A Signal-Guided Mixture-of-Experts Ensemble Architecture

Traditional ensemble machine learning methods typically combine predictions via uniform aggregation or simpleweighted averaging, often lacking awareness of the structural geometry and local complexity of input data. This paper introduces the Track/Rail Algorithm (TRA), a novel Mixture-ofExperts (MoE) architecture that integrates Switch Transformerinspired routing with a signal-guided expert gating mechanism. Unlike conventional MoE models that rely solely on raw input features for routing decisions, TRA incorporates a Signal Extraction Layer that derives five structural meta-features:expert disagreement, prediction entropy, feature density, cluster distance, and anomaly scores. These signals dynamically guidethe router to allocate samples to a heterogeneous pool of expert tracks-including Random Forest, XGBoost, LightGBM, andSupport Vector Machines-each specialized through K-Meansbased data partitioning. The architecture is further enhancedby temperature-scaled soft routing, dynamic track spawning for high-uncertainty regions, and a residual correction mechanism(TRA-Boost). An extensive empirical evaluation comprising 1,900 total evaluations across 21 unique datasets and 10 distinctmodels was conducted. Experimental results demonstrate that TRA achieves highly competitive performance, attaining perfectaccuracy (1.000) on the Iris dataset and the lowest RMSE (53.16) on the Diabetes regression task. While gradient boostingframeworks such as CatBoost and LightGBM remain strong baselines, TRA demonstrates superior adaptability in specificstructural domains. These results suggest that incorporating model-consensus and data-geometry signals into the routingprocess significantly enhances expert specialization and ensemble robustness.