Accurate spatiotemporal traffic flow forecasting is significant for the modern traffic management and control. In order to capture the spatiotemporal characteristics of the traffic flow simultaneously, we propose a novel spatiotemporal residual graph attention network (STRGAT). First, the network adopts a deep full residual graph attention block, which performs a dynamic aggregation of spatial features regarding the node information of the traffic network. Second, a sequence-to-sequence block is designed to capture the temporal dependence in the traffic flow. The traffic flow data with weekly periodic dependencies are also integrated and STRGAT is used for traffic forecasting of traffic road networks. The experiments are conducted on three real data sets in California, USA. Results verify that our proposed STRGAT is able to learn the spatiotemporal correlation of traffic flow well and outperforms the state-of-the-art methods.

Time-domain analysis and frequency-domain analysis are two angles of view for us to study and survey a continuous function. We observe the function approximation problems from the frequency domain. We consider that a single-frequency sine function can be approximated by two Gaussian kernels in one period. According to that, we present that the first maximum amplitudes as well as their frequencies and initial phases can be used to determine the initial number, centers and widths of radial basis function (RBF) kernels. After the initial structure of an RBF network is determined like that, a small number of RBF kernels can be added in order to further improve the local approximation accuracy. The above viewpoint is verified by two approximation examples.

This paper proposes a task decomposition method, which divides a large-scale learning problem into multiple limited-scale pairs of training subsets and cross validation (CV) subsets. Correspondingly, modular multilayer perceptrons are set up. At first, one training subset only consists of its own class and several most neighboring categories, and then some classes in the CV subset are moved into it according to the generalization error of the module. This work presents an empirical formula for selecting the initial number of hidden nodes, and a method for determining the optimal number of hidden units with the help of singular value decomposition. The result for letter recognition shows that the above methods are quite effective.

This paper proposes a modular RBF-type network classifier consisting of multiple single-layer RBF networks and perceptrons for solving large-sample and high-dimensional classification problems. The method for optimally determining the number, locations and widths of RBF kernels and the target values of Gaussian activation functions is gone into details. The presented method, which only propagates errors one layer backwards, has much lower computational complexity than the backpropogation algorithm used in multilayer feedforward perceptrons and RBF networks. The result for letter recognition shows that modular RBF-type network classifiers as well as the adaptively learning algorithm have faster convergence rate, higher classification accuracy, larger probability to get optimal structures, and better performance to reach global minimum points, compared with the standard RBF networks and multilayer feedforward perceptrons.