Predicting chaotic time series and replicating chaotic attractors based on two novel echo state network models
Introduction
As an important role of nonlinear science research, chaos phenomenon has extremely important implications and application value. Chaotic systems exhibit very complex and interesting dynamical behaviors, such as chaotic attractors, Hopf branches, period-doubling branches and homoclinic orbits. Although chaos has been known for nearly half a century, it is a pity that there is still no universally accepted and strict mathematical definition of chaos until today. Nevertheless, the general characteristics of chaos have been summarized: the sensitive dependence on initial value, long-term unpredictability, Lyapunov index greater than 0 etc. The Lorenz system [1] is the first physical and mathematical model of chaos in the history of chaos evolution, which is an important milestone. Over the years, some more complex and diverse chaotic systems have been artificially constructed, like Rössler system [2], Chua system [3], Chen system [4].
With the rapid development of artificial intelligence today, the application of artificial intelligence to deal with chaos related problems has become a topic of great concern in dynamics. The problems associated with chaos are prediction of chaotic time series, reconstruction of chaotic attractor, chaos control. The use of artificial intelligence tools for example neural networks is one of the alternatives in this field that is increasingly used by researchers in different engineering disciplines. Artificial neural network (ANN) is utilized to simulate the structure and function of biological neural network [5]. Owing to the fast-growing contemporary neurobiology, mathematics, physics, and computer science, a large number of artificial neural network models have been proposed, and the research has gradually become a focus [6], [7], [8], such as the McCullochPitts neuron model [9], linear threshold function neuron model [10], [11], backpropagation neural network model [12], [13], [14], recurrent neural network (RNN) model [15], [16], [17].
The reason why RNN is more suitable for human brain simulation is that RNN has a better ability to describe the dynamic properties of the system, compared with the feed forward neural network [18]. Under normal circumstances, RNNs can infinitely approach any complex nonlinear dynamic system. These advantages make RNN widely used in prediction [19], nonlinear system identification [20], and adaptive filtering [21]. Nonetheless, some inevitable problems still exist in RNNs, such as, the network structure and training algorithm are complex, the computation is large, the convergence speed is slow, and the error gradient may disappear or produce distortion with the increase of training steps. Some improvements have been achieved in recent years [22], [23], [24]. Gradient vanishing phenomenon is common but difficult to overcome in neural network training. Nonetheless, with the further development of artificial intelligence, experts have proposed some solutions to the gradient vanishing problem. In 1998, Hochreiter theoretically analyzed the gradient vanishing problem of recurrent neural networks, but the improved algorithm also has some drawbacks (e.g. practicable only for discrete problems) [25]. Hinton mentioned in 2006 that in order to improve the problem of gradient vanishing, unsupervised layer-by-layer training is adopted. After the pre-training layer-by-layer is complete, the entire network is “fine-tuned”. This method has certain benefits, but it is not used much now [26]. In recent years, a new multilayer neural network was proposed that adds linear neurons, which is easier to train than traditional MLP networks and reduces the effects of gradient vanishing [27]. Recurrent identity network (RIN), which allows a plain recurrent network to overcome the gradient vanishing problem while training very deep models without the use of gates, was proposed in 2019. On multiple sequence modeling benchmarks, the RIN model demonstrates competitive performance and converge faster in all tasks [28].
As mentioned above, gradient disappearance brings indelible problems to the training of neural networks. However, Jaeger proposed ESN in 2001 [29], and the training process of this model is greatly different from that of traditional RNN. Training ESN model does not need to use gradient to update parameters, but use the state variables sampled during network training and linear output to get the optimal parameters. So ESN model, in a sense, avoids the problem of local minima and gradient disappearance. The merits of ESN are that the process of training considerably reduced the amount of computation, overcame the problem of memory fading, and also solved the structural defects of the traditional RNN. The most significant difference between ESN and traditional RNNs lies in the structure of reservoir layer. In 2002, ESN and liquid state machine (LSM) are collectively referred to as reservoir computing [30]. The weights of each layer in traditional RNNs are computed by the gradient descent algorithm. While in ESN, only the weights between reservoir layer and output layer need to be trained, which greatly reduces the calculation amount and time of the traditional RNN and makes the calculation efficiency higher. Gradually, based on these unique advantages, ESN has attracted increasing attention in the field of neural networks [31], [32]. Not long after ESN model was proposed, it has been employed to chaotic time series prediction. For the reconstruction attractor, notably, in 2017, Pathak, Lu et al. used ESN methods to replicate spatiotemporal chaotic attractors and calculate the Lyapunov exponents from data [33]. The paper caused a stir in the field of chaos and was immediately reported by the major media. For the time series data of actual production and life, the novel ESN were proposed, the calculation amount are greatly reduced and the prediction accuracy was improved for the task of wind speed and wind direction forecasting [34]. In addition, ESN has also been applied in the field of handwritten digit recognition [35].
In this paper, two novel ESN models are proposed based on homotopy and combination activation function, and successfully applied to chaotic time series. The rest of this paper is organized as follows. In Section 2, the Leakage Integral Echo State Network (LIESN), the proposed Homotopy Activation Function Leakage Integral Echo State Network (HAF-LIESN) and Combinatorial Activation Function Leakage Integral Echo State Network (CAF-LIESN) models are briefly introduced. Moreover, how to train the output weight is given by a detailed derivation process. We apply the proposed models to the prediction of MG chaotic time series and Rössler chaotic time series in Section 3. Meanwhile, the evaluate methods are mean squared error (MSE), root mean squared error (RMSE) and measures average error (MAE). Several parameters play key roles in network training, such as spectral radius, sparse degree etc, and Section 3 also investigates their effects on network performance. What is more, the Rössler chaotic attractor is replicated. In view of the two models proposed in this paper, the parameter optimization of the model is discussed in Section 3.3, which leads to better network performance. The final section summarizes our contributions and gives some suggestions for future works.
Section snippets
A brief overview of LIESN, HAF-LIESN, CAF-LIESN and output weight
ESN is a variant of RNN neural network, proposed by Jaeger in 2001 [29], [36]. Only the network output weight is trained by linear regression, whereas the connections of input-to-reservoir and reservoir to-reservoir are fixed after random initialization. In this way, the computational complexity of the training process is reduced considerably.
Results
In this section, two basic problems are employed to prove the effectiveness of the proposed models, namely, the Mackey-Glass (MG) time series and Rössler attractor. Each dataset is divided into two parts for training and testing, and the length is denoted as and , respectively.
To evaluate the experimental results, MSE, RMSE and MAE were used to assessed the performance of the proposed methods. MSE measures mean square error, which is closer to zero is better. Taking the square root
Conclusion
In this paper, based on the idea of homotopy and combinatorial activation functions, HAF-LIESN and CAF-LIESN are proposed, successfully applied to chaotic time series and the replication of chaotic attractor. The datasets of chaotic time series are MG and Rössler attractors respectively. According to the results, the prediction error of HAF-LIESN and CAF-LIESN models are smaller than other ESNs based on the relevant evaluation criteria (MSE, RMSE, MAE), which the effectiveness of the activation
Funding
This research is supported by National Basic Research Program of China (Grant No. 2013CB834100), National Natural Science Foundation of China(Grant No. 12071175), National Natural Science Foundation of China(Grant No. 11571065) and JilinDRC(Grant No. 2017C028-1).
CRediT authorship contribution statement
Yuting Li: Conceptualization, Methodology, Data curation, Visualization, Investigation, Formal analysis, Writing – original draft. Yong Li: Conceptualization, Funding acquisition, Methodology, Resources, Supervision, Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
Thanks to Professor Yong Li for his suggestions and guidance in this work.
Yuting Li, a Ph.D. candidate, College of Mathematics, Jilin University. Her research focuses on prediction of chaotic time series, machine learning, and deep learning applications to dynamics.
References (41)
- et al.
Exponential stability analysis of quaternion-valued neural networks with proportional delays and linear threshold neurons: Continuous-time and discrete-time cases
Neurocomputing
(2020) - et al.
Topographic property of backpropagation artificial neural network: From human functional connectivity network to artificial neural network
Neurocomputing
(2020) - et al.
A new hybrid quadratic regression and cascade forward backpropagation neural network
Neurocomputing
(2016) - et al.
Identification and optimal control of nonlinear systems using recurrent neural networks and reinforcement learning: An overview
Neurocomputing
(2021) - et al.
International journal of electrical power with energy systems
Proc. IEEE Int. Symp. Ind. Electron.
(2014) - et al.
An effective multivariate time series classification approach using echo state network and adaptive differential evolution algorithm
Expert Syst. Appl.
(2016) - et al.
Wind speed and wind direction forecasting using echo state network with nonlinear functions
Renew. Energy
(2019) Deterministic non-periodic flows
Atoms. Sci.
(1963)Continuous chaos four prototype equations
Ann. N. Y. Acad. Sci.
(1979)- et al.
The double scroll family. Part I:Rigorous proof of chaos
IEEE Trans. Circuits. Syst.
(1986)
Yet another chatic attractor
Int. J. Bifur. Chaos
Multi-instant observer design of discrete-time fuzzy systems: A ranking-based switching approach
IEEE Trans. Fuzzy Syst.
Finite-horizon Η ∞ state estimation for time-varying neural networks with periodic inner coupling and measurements scheduling
IEEE Trans. Syst. Man Cybern. Syst.
Remote estimator design for time-Delay neural networks using communication state information
IEEE Trans. Neural Netw. Learn. Syst.
A logical calculus of the ideas immanent in nervous activity
Self-organized formation of topologically correct feature maps
Biol. Cybern.
Learning representations by back-propagating errors
Nature
Neural networks and physical systems with emergent collective computational abilities
Proc. Natl. Acad. Sci.
Cited by (7)
Short-term traffic flow prediction based on secondary hybrid decomposition and deep echo state networks
2023, Physica A: Statistical Mechanics and its ApplicationsAnalysis of the configuration characteristics of the chaotic attractor for machine tool dynamic error under the singular spectrum
2023, Journal of Manufacturing ProcessesDetection of Weak Pulse Signal in Chaotic Noise - Based on Att-CNN-LSTM model
2024, Research SquareQuantum image chaos encryption scheme based on quantum long-short term memory network
2023, Wuli Xuebao/Acta Physica SinicaChaotic Time Series Prediction of Multi-Dimensional Transformer Based on Bionic Reconfigurable Three-Dimensional Scanning Robot
2023, 2023 IEEE International Conference on Mechatronics and Automation, ICMA 2023
Yuting Li, a Ph.D. candidate, College of Mathematics, Jilin University. Her research focuses on prediction of chaotic time series, machine learning, and deep learning applications to dynamics.
Yong Li received the B.S. degree in Mathematics from the Northeast Normal University, Changchun, China, in 1982, the M.S. degree in Mathematics from the Jilin University Changchun, China, in 1985, and the Ph.D. degree in Mathematics from the Jilin University Changchun, China, in 1990. He now is the Director of the Institute of Mathematics, Jilin University. His research focuses on differential equations and dynamical systems, and he has published numerous papers in well-recognized journals including Math. Ann., Trans. Amer. Math. Soc., Arch. Ration. Mech. Anal., Ann. Henri Poincaré, J. Nonlinear Sci., J. Differential Equations etc. He is an editor of Journal of Nonlinear Mathematical Physics and Electronic Research Archive. He received the National Science Fund for Distinguished Young Scholars in 2002 and the second prize in China’s State Natural Science Award in 2016.