Neural Network-Based Online Model Identification and Nonlinear Adaptive Predictive Control of Self-Balancing Two-Wheel LEGO Mindstorms NXTway-GS Robot

The main challenge of the well-celebrated Levenberg-Marquardt algorithm (LMA) is the selection of the searching direction and adaptation parameters. Secondly, the implementation of the LMA for online model identification has faced challenges as it is a batch optimization. As a third challenge, the solution of the Levenberg-Marquardt based on the full-Newton nonlinear optimization (FNNO) for online applications have been limited due to its unguaranteed positive definiteness. This paper presents two versions of the modified Levenberg-Marquardt algorithm (MLMA) for neural network model identification and adaptive predictive control for the online dynamic model identification and adaptive control of a self-balancing two-wheel LEGO Mindstorms NXTway-GS robot. The first version is the online-window-approach of the modified Levenberg-Marquardt algorithm (OWA-MLMA) based on approximate Guass-Newton algorithm (AGNA) for training neural network model predictor. The second version is a neural network-based adaptive predictive control (APC) algorithm based on the full-Newton nonlinear optimization of the modified Levenberg-Marquardt algorithm (FNNO-MLMA) for online adaptive control. A NNARMAX model predictor for the NXT robot is first trained and validated using the OWA-MLMA based on AGNA. The validated NNRAMAX model is then used for the design of the NN-APC based on FNNO-MLMA. Finally, the model identification based on OWA-MLMA and APC based on FNNO-MLMA schemes are simulated in closed-loop for online NNARMAX model identification and adaptive predictive control of the NXT robot. The comparison of the proposed OWA-MLMA based on AGNA shows superior performance over the recursive incremental back-propagation algorithm (INCBPA) while the proposed NN-APC based on FNNO-MLMA shows excellent control and good tracking performances over a discrete-time fixed-parameter proportional-integral-derivative (PID) controller for the NXT robot control. The simulation results show that the developed OWA-MLMA based on AGNA and the APC based on FNNO-MLMA can be adapted for the online dynamic modeling, automation, control and robotics applications.