Reservoir computing-broad learning system with Bayesian Optimization for robotic arm motion trajectory prediction

Deep learning models perform well in time series prediction tasks. However, these methods generally have issues such as high demand for computing resources and a tendency to over-fit. In this context, the broad learning system (BLS) emerged. Due to its unique flat structure, BLS can be constructed rapidly and is less likely to over-fit. However, the BLS lacks a feedback structure, which poses certain limitations when it comes to handling time series prediction tasks. To solve this problem, we use reservoir computing (RC) to replace the static enhancement nodes in the enhancement layer of BLS, resulting in the RC-BLS model. This model is more suitable for handling time series data. In addition, we optimize the hyperparameters of the RC-BLS model using the Bayesian Optimization algorithm (BO). The application objects of this paper are four real-world time series datasets. Three of them are public datasets, and the other is a custom dataset collected from a precision robotic arm in a human-robot collaboration (HRC) environment. Comparison experiments with several different machine learning algorithms show that the BO-RC-BLS model has more minor prediction errors and better performance in real-world time series prediction. This innovative approach not only solves the time-consuming problem of manual hyperparameter tuning but also improves the performance of the model.