Dynamic environment prediction on unmanned mobile manipulator robot via ensemble convolutional randomization networks

Abstract

This paper aims to improve the prediction for dynamic environments on unmanned mobile manipulator robot system. When it comes to a simple prediction task, the mobile manipulator robot is undesirable to spend lots of time training a complex neural network. In this paper, the structure of randomization-based neural networks with the convolutional layer named conv-RNet is tested, which reveals that conv-RNet seems to be better than randomization-based neural network with the fully connected layer in terms of accuracy, parameters and computational complexity. Based on the conv-RNet, an ensemble learning architecture named Ensemble convolutional randomization-based neural networks (EC-RNet) is proposed to further optimize the network performance. Here, three problems are mainly solved to optimize this ensemble architecture. The first is how to extract local feature information and decrease the computational complexity. The convolutional layer is used to filter the input to extract features. For producing the fixed number of the hidden layer nodes, the convolutional operation could extract more abundant feature information while maintaining few parameters and low computational complexity than a fully connected operation. The second is how to calculate the number of component learners. Little probability event is used to build a relationship between the number of component learners and the data number in each component dataset. For a given component data, the minimum number of component learners will be determined by this relationship. The third is how to combine the results of component learners to obtain the final result. The confidence level is introduced as the weight to measure the relationship between component results and final results. The combination of component results attaching to different confidence levels is used to calculate the final result. On CIFAR-10, MNIST handwritten digits dataset, and UCI dataset, EC-RNet achieves high accuracy, low computational complexity, and few parameters. Moreover, Experiments on above three datasets show that the proposed EC-RNet+RVFL structure outperforms the proposed EC-RNet+ELM structure. In the real world, the EC-RNet is deployed on the mobile robot and achieves more reliable performance.