Optimization of Training Data Set Based on Linear Systematic Sampling to Solve the Inverse Kinematics of 6 DOF Robotic Arm with Artificial Neural Networks

The amount of data that can be represented in the workspace of a robotic manipulator can be a factor that has a decisive influence on the processing time and that ensures the success of the knowledge extraction algorithms. In this study, two data sets were generated by analyzing the direct kinematics of a six-degree-of-freedom robotic manipulator. The first set was generated with a size greater than 4 billion data and the second set with a quantity greater than 350 thousand data. To solve the data volume problem, a data reduction filtering algorithm based on the linear systematic sampling technique was implemented. To validate the filtering algorithm, the training of two neural network architectures was performed, measuring the performance and generalizability in both networks due to the application of the filter on the data. The network architectures used were a back propagation neural network and a generalized regression neural network. For the first network, the optimal parameters were determined by applying a robust design methodology based on the Taguchi philosophy applied to the design of neural networks. For the second, a comparative performance model was used to determine the best constant propagation value for network training. In the results obtained, an increase in the generalizability was observed when using the data set previously treated by the filter in both network architectures. In the testing stage, a chi-square statistical analysis of less than 5% was considered to validate the application of the filtering algorithm, managing to maintain a prediction of 83% of the test data within the same margin of error.