The Reservoir Computing Research For Echo State Network And Conceptor Network

Echo state network(ESN) is a typical reservoir computing(RC) model, mapping low dimensional information to the high dimensional state space, and converting high dimensional information to low dimensional output via a simple linear regression learning algorithm. The state space with high dimension is called reservoir, which greatly reduces the computational complexity of the traditional artificial neural network and overcome memory decay problems. As a key processing units, the input coding mode of reservoir directly affect the reading precision of the output layer. Hence, reservoir model optimization is always a research focus in the field. There have been many complex network models. Among them, the cortex multi-cluster network has attracted significant attention because of its rich dynamics and bionic characteristics. But randomness existing in the network algorithm could lead to network computing capacity exist a large fluctuation, and the algorithm parameters are difficult to adjust. So, reservoir model needs the further optimization. In addition, the conceptor network, is a newly proposed reservoir computing model, and its dynamics module has a good scalability. But at present, the network structure design is single, only using the traditional random network, which has strong node coupling and limited computing capacity. Consequently, it is very important to optimize reservoir structure for conceptor network.For ESN, we proposed two reservoir optimal models. On the one hand, based on cortex multi-cluster structure, we proposed a prior-data oriented cluster reservoir. We adjusted reservoir topology off-line using prior data samples to make it more adapt to a kind of computing tasks. By the Mackey-Glass series prediction experiments, we can discover that the prior-data oriented cluster reservoir significantly increased computing accuracy comparing with traditional random network and cortex cluster network, and it has higher structure complexity and small-world properties.On the other hand, combining neuronal intrinsic plasticity(IP) with the cluster reservoir structure, we proposed intrinsic plasticity cluster reservoir. We analyzed two kinds of IP rules‘ affects for neuronal input response, and demonstrated their effectiveness via NARMA series prediction examples: the cluster reservoirs with IP broke through the cluster reservoir‘s precision bottleneck caused by the randomness in the process of network generating, and greatly improve the prediction accuracy of random reservoir cortex cluster reservoir. Furthermore, the cluster reservoir with Li‘s IP rule has the more prominent precision advantages.For conceptor network, based on the researching for complex network, this paper proposed a complex network reservoir based on the Lorentz’s time series phase space reconstruction. We analyzed the network structure characteristics, and compared its reconstructing capability for input patterns with random network and cortex cluster network. The experiments show that the phase space reconstruction network has less error for signal reconfigurable computing. Furthermore, under the condition of changing load pattern‘s number, it still keep this advantages.