Abstract
Training of recurrent neural networks (RNNs) introduces considerable computational complexities due to the need for gradient evaluations. How to get fast convergence speed and low computational complexity remains a challenging and open topic. Besides, the transient response of learning process of RNNs is a critical issue, especially for online applications. Conventional RNN training algorithms such as the backpropagation through time and real-time recurrent learning have not adequately satisfied these requirements because they often suffer from slow convergence speed. If a large learning rate is chosen to improve performance, the training process may become unstable in terms of weight divergence. In this paper, a novel training algorithm of RNN, named robust recurrent simultaneous perturbation stochastic approximation (RRSPSA), is developed with a specially designed recurrent hybrid adaptive parameter and adaptive learning rates. RRSPSA is a powerful novel twin-engine simultaneous perturbation stochastic approximation (SPSA) type of RNN training algorithm. It utilizes three specially designed adaptive parameters to maximize training speed for a recurrent training signal while exhibiting certain weight convergence properties with only two objective function measurements as the original SPSA algorithm. The RRSPSA is proved with guaranteed weight convergence and system stability in the sense of Lyapunov function. Computer simulations were carried out to demonstrate applicability of the theoretical results.
Similar content being viewed by others
Abbreviations
- k :
-
Discrete time step
- m :
-
Output dimension
- n I :
-
Input dimension
- n h :
-
Hidden neuron number
- p v :
-
= m × n h . Dimension of output layer weight vector
- p w :
-
= n i × n h . Dimension of hidden layer weight vector
- l :
-
Number of time-delayed outputs
- u i (k):
-
\(i=1,\ldots,m.\) External input at time step k
- x(k):
-
\(\in R^{n_{I}}.\) Neural network input vector
- \(\Updelta x(k)\) :
-
\(\in R^{n_{I}}.\) Perturbation vectors on input x(k − 1)
- y(k):
-
\(\in R^{m}.\) Desired output vector
- \(\hat{y}(k)\) :
-
\(\in R^{m}.\) Estimated output vector
- \(\varepsilon(k)\) :
-
\(\in R^{m}.\) Total disturbance vector
- e(k):
-
\(\in R^{m}.\) Output estimation error vector
- \(\overline{V}(k)\) :
-
\(\in R^{p^{v}}.\) Estimated weight vector of output layer
- \(\overline{V}^{*}(k)\) :
-
\(\in R^{p^{v}}.\) Ideal weight vector of output layer
- \(\tilde{\overline{V}}(k)\) :
-
\(\in R^{p^{v}}.\) Estimated error vector of output layer
- \(\overline{W}(k)\) :
-
\(\in R^{p^{w}}.\) Estimated weight vector of hidden layer
- \(\overline{W}^{*}(k)\) :
-
\(\in R^{p^{w}}.\) Ideal weight vector of hidden layer
- \(\tilde{\overline{W}}(k)\) :
-
\(\in R^{p^{w}}.\) Estimated weight vector of hidden layer
- \(\hat{W}_{j,:}(k)\) :
-
\(\in R^{n_{i}}.\) The jth row vector of hidden layer weight matrix \(\hat{W}(k)\)
- δ v(k):
-
Equivalent approximation errors of the loss function of output layer
- δ w(k):
-
Equivalent approximation errors of the loss function of hidden layer
- c :
-
Perturbation gain parameter of the SPSA
- \(\Updelta^{v}\) :
-
\(\in R^{p^v}.\) Perturbation vectors of output layer
- r v :
-
\(\in R^{p^v}.\) Perturbation vectors of output layer
- \(\Updelta^{w}\) :
-
\(\in R^{p^w}.\) Perturbation vectors of hidden layer
- r w :
-
\(\in R^{p^w}.\) Perturbation vectors of hidden layer
- \(H(\overline{W}(k),x(k))\) :
-
\(=H(k)\in R^{m\times p^w}.\) Nonlinear activation function matrix
- h j (k):
-
\(=h\left(\hat{W}_{j,:}(k)x(k)\right).\) The nonlinear activation function and the scalar element of \(H(\overline{V}(k),x(k))\)
- α v :
-
Adaptive learning rate of output layer
- α w :
-
Adaptive learning rate of hidden layer
- α v, α w :
-
Positive scalars
- ρ v :
-
Normalization factor of output layer
- ρ w :
-
Normalization factor of hidden layer
- β v :
-
Recurrent hybrid adaptive parameter of output layer
- β w :
-
Recurrent hybrid adaptive parameter of hidden layer
- μ j (k):
-
\(\in R;1\leq j\leq n_{h}.\) Mean value of the input vectors of the jth hidden layer neuron
- \(\tilde{\mu}_{j}(k)\) :
-
\(\in R;1\leq j\leq n_{h}.\) Mean value of the hidden layer weight vector of the jth hidden layer neuron
- τ :
-
Positive scalar
- λ :
-
Positive gain parameter of the threshold function
- η :
-
A small perturbation parameter
References
Hopfield JJ (1982) Neural networks and physical systems with emergent collective computational abilities. Proc Natl Acad Sci 79:2554–2558
Talebi HA (2009) A recurrent neural-network-based sensor and actuator fault detection and isolation for nonlinear systems with application to the satellite’s attitude control subsystem. IEEE Trans Neural Netw 20:45–60
Hou ZG, Gupta MM, Nikiforuk PN, Tan M, Cheng L (2007) A recurrent neural network for hierarchical control of interconnected dynamic systems. IEEE Trans Neural Netw 18:466–481
Al Seyab RK, Cao Y (2008) Nonlinear system identification for predictive control using continuous time recurrent neural networks and automatic differentiation. J Process Control 18:568–581
Song Q, Xiao J, Soh YC (1999) Robust backpropagation training algorithm for multilayered neural tracking controller. IEEE Trans Neural Netw 10:1133–1141
Song Q, Wu Y, Soh YC (2008) Robust adaptive gradient-descent training algorithm for recurrent neural networks in discrete time domain. IEEE Trans Neural Netw 19:1841–1853
Song Q, Spall JC, Soh YC, Ni J (2008) Robust neural network tracking controller using simultaneous perturbation stochastic approximation. IEEE Trans Neural Netw 19:817–835
Song Q (2008) On the weight convergence of Elman networks. IEEE Trans Neural Netw 21:463–480
Haykin S (1999) Neural networks: a comprehensive foundation. Printice Hall, New Jersey
Mandic DP, Chambers JA (2001) Recurrent neural networks for prediction: learning algorithms, architectures and stability. Wiley, New York
Spall JC (1992) Multivariate stochastic approximation using a simultaneous perturbation gradient approximation. IEEE Trans Autom Control 37:332–341
Maeda Y, Wakamura M (2005) Simultaneous perturbation learning rule for recurrent neural networks and its FPGA implementation. IEEE Trans Neural Netw 16:1664–1672
Trunov AB, Polycarpou MM (2000) Automated fault diagnosis in nonlinear multivariable systems using a learning methodology. IEEE Trans Neural Netw 11:91–101
Spall JC, Cristion JA (1998) Model-free control of nonlinear stochastic systems with discrete-time measurements. IEEE Trans Autom Control 43:1198–1210
Spall JC, Cristion JA (1997) A neural network controller for systems with unmodeled dynamics with applications to wastewater treatment. IEEE Trans Syst Man Cybern Part B Cybern 27:369–375
Maeda Y, De Figueiredo RJP (1997) Learning rules for neuro-controller via simultaneous perturbation. IEEE Trans Neural Netw 8:1119–1130
Werbos PJ (1988) Generalization of backpropagation with application to a recurrent gas market model. Neural Netw 1:339–356
Rumelhart D, Hinton G, Williams R (1986) Learning internal representations by error backpropagation. Parallel Distrib Process 1:318–362
Williams RJ, Zipser D (1989) A learning algorithm for continually running fully recurrent neural networks. Neural Comput 1:270–280
Williams RJ, Zipser D (1995) Gradient-based learning algorithms for recurrent networks and their computational complexity. Backpropag Theory Archit Appl 2:433–501
Lin T, Giles C, Horne B, Kung S (1997) A delay damage model selection algorithm for NARX neural networks. IEEE Trans Signal Process 45:2719–2730
Park Y, Murray T, Chen C (2002) Predicting sun spots using a layered perceptron neural network. IEEE Trans Neural Netw 7:501–505
Ljung L (2010) Perspectives on system identification. Annu Rev Control 34(1):1–12
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Xu, Z., Song, Q. & Wang, D. A robust recurrent simultaneous perturbation stochastic approximation training algorithm for recurrent neural networks. Neural Comput & Applic 24, 1851–1866 (2014). https://doi.org/10.1007/s00521-013-1436-5
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00521-013-1436-5