Discovering Stick-Slip-Resistant Servo Control Algorithm Using Genetic Programming

The stick-slip is one of negative phenomena caused by friction in servo systems. It is a consequence of complicated nonlinear friction characteristics, especially the so-called Stribeck effect. Much research has been done on control algorithms suppressing the stick-slip, but no simple solution has been found. In this work, a new approach is proposed based on genetic programming. The genetic programming is a machine learning technique constructing symbolic representation of programs or expressions by evolutionary process. In this way, the servo control algorithm optimally suppressing the stick-slip is discovered. The GP training is conducted on a simulated servo system, as the experiments would last too long in real-time. The feedback for the control algorithm is based on the sensors of position, velocity and acceleration. Variants with full and reduced sensor sets are considered. Ideal and quantized position measurements are also analyzed. The results reveal that the genetic programming can successfully discover a control algorithm effectively suppressing the stick-slip. However, it is not an easy task and relatively large size of population and a big number of generations are required. Real measurement results in worse control quality. Acceleration feedback has no apparent impact on the algorithms performance, while velocity feedback is important.

Multi-Gene Genetic Programming-Based Identification of a Dynamic Prediction Model of an Overhead Traveling Crane

This paper proposes a multi-gene genetic programming (MGGP) approach to identifying the dynamic prediction model for an overhead crane. The proposed method does not rely on expert knowledge of the system and therefore does not require a compromise between accuracy and complex, time-consuming modeling of nonlinear dynamics. MGGP is a multi-objective optimization problem, and both the mean square error (MSE) over the entire prediction horizon as well as the function complexity are minimized. In order to minimize the MSE an initial estimate of the gene weights is obtained by using the least squares approach, after which the Levenberg–Marquardt algorithm is used to find the local optimum for a k-step ahead predictor. The method was tested on both a simulation model obtained from the Euler–Lagrange equation with friction and the experimental stand. The simulation and the experimental stand were trained with varying control inputs, rope lengths and payload masses. The resulting predictor model was then validated on a testing set, and the results show the effectiveness of the proposed method.