Evaluation of Artificial Neurocomputing Algorithms and Their Metacognitive Robustness in Predictive Modeling of Fuel Consumption Rates During Tillage

Abstract

Artificial intelligence (AI) requires complex neurocomputing algorithms (NCAs) for robustness in predictive modeling. Since 1990, a plethora of NCAs have evolved, challenging selection of the most intelligent and error-tolerant model for predicting fuel consumption rate (ØFuel) in tillage. This study adopted 12 recent NCAs in modeling 72 neurocomputing architectures and evaluated their metacognitive robustness in predicting ØFuel using 80 tillage datasets. These included Levenberg-Marquardt (trainlm), Quasi-newton (trainbfg), Powell-Beale conjugate gradient (traincgb), Scaled conjugate gradient (trainscg), Fletcher-reeves conjugate gradient (traincgf), Polak-ribiére conjugate gradient (traincgp), Onestep secant (trainoss), Bayesian regularization (trainbr), Resilient backpropagation (trainrp), Gradient descent (traingd), Learning rate gradient descent (traingdx) and Gradient descent momentum (traingdm). Tillage variables were sequentially subjected to NCA architectures learning on logsig, Purelin, and tansig neuro-activation functions, targeting 1000 epochs. Neuro-cognitive performance was evaluated through broad multi-criteria of regressed neuron input–output functions using heuristic and accuracy metrics, i.e., training time (t), number of hidden layers (NL), and neurons (Nn), optimal epoch (Eopt), Coefficient of correlation (R) and determination (R2), Mean square error (MSE), Root mean square error (RMSE), Mean absolute error (MAE), Mean absolute percentage error (MAPE), Sum square error (SSE), Prediction scatter (Tscatter), Coefficient of variation (CV) and Prediction accuracy (PA). Prediction reliability and robustness were evaluated using a20-index (a20), Willmott’s index of agreement (IOA), Index of scatter (IOS), Variance accounted for (VAF), Performance index (PI), Total score rank (ξtotal), Sensitivity index (SAij) and Anderson-Darling (AD) test. Results indicated that varying NL and Nn did not correspond with t, Eopt and PA but only occurred in single-layered trainbr and double-layered trainlm. Although PA increased at reduced convergence error limits, relationships among Eopt, t, and PA were unclear. Single-layered trainbr (14-20-1) outperformed all NCAs with R (0.9993), R2 (0.9986), MSE (6.34e-06), RMSE (0.0025), MAE (0.003), MAPE (1.4e-06), SSE (8e-04), R2All (0.9998) Tvalue (1.0), CV (3.232 %) and PA (99.81 %). However, trainlm (14-20-10-1) most accurately modeled double-layered architectures i.e. R (0.9932), R2 (0.9864), MSE (7.4e-05), RMSE (0.009), MAE (0.004), MAPE (0.003), SSE (0.003), R2All (0.9929), Tvalue (0.9998), CV (3.212 %) and PA (97.8 %). Nonetheless, trainbr (14-20-1) achieved the best reliability metrics of a20 (100 %), VAF (99.98 %), IOA (1.00), PI (1.9971), IOS (0.0115), ξtotal (284) and AD-test (2.6852). Sensitivity analysis revealed that tractive force had the most significant influence on ØFuel while tire inflation pressure had the least. Learning on transig neuro-activation function in single-layered neural network, metacognitive robustness of trainbr is superiorly merited for predicting ØFuel during tillage.