Improved recurrent neural networks for solving Moore-Penrose inverse of real-time full-rank matrix
Introduction
It is known that Moore-Penrose inverse (i.e., pseudoinverse) plays a key role in science and engineering, e.g., system modeling and design in a variety of applications involving robotics [1], signal processing [2], image restoration [3] and pattern recognition [4]. Therefore, how to efficiently compute Moore-Penrose inverse of a matrix has become a critical problem in real-time applications. Although there exist many numerical methods for computing Moore-Penrose inverse such as singular value decomposition (SVD) [5], Greville’s recursive method [6], Newton iteration method [7] and so on, and from the viewpoint of computation, their serial computational schemes are inefficient when they are applied to large-scale and real-time problems [8], [9]. In fact, the computational costs of these known methods increase significantly with the increase of the size of the matrices because of the complexity being usually required, which makes the solving process time-consuming. However, there are real-time matrix Moore-Penrose inversion problems in practice, for example, the inverse kinematic problem for online control of redundant robot manipulator needs the online solution of Moore-Penrose inverse [10], [11] (see also the Section 6 in this paper). Obviously, the existing serial algorithms are not suitable for the online solution of Moore-Penrose inverse of a matrix. For this reason, the parallel computing methods become the natural choices for reducing computational costs and increasing the computational efficiency, especially for the online solution of Moore-Penrose inverse.
Due to the parallel nature and convenience of hardware implementation, recurrent neural network (RNN), originated from Hopfield neural network, has been deemed as a powerful tool to solve online matrices inversion problems [12], [13], [7], [14], [15], [16]. Over the past decades, many recurrent neural network models for computing Moore-Penrose inverse of a full-rank matrix have been proposed and investigated [17], [18], [11], [19]. As a special recurrent neural network, gradient neural network (GNN) is designed intrinsically for solving static problems, and it has become an effective tool for solving the Moore-Penrose inverse of a constant matrix [18], [17], [20]. Nevertheless, many methods including GNN models designed for static problems may not be effective for solving time-varying ones [10], [9]. To overcome this shortcoming, Zhang neural network (ZNN), as another special case of RNN, has been proposed for both time-invariant and time-varying problems solving [21], [22], [23]. Meanwhile, ZNN models have been successfully applied to solve matrix Moore-Penrose inversion [10], [24], [25], [26]. Recently, based on the GNN and ZNN models, Lv et al. [8] presented two more efficient RNN models which were named as novel neural network (NNN) models. By the detailed analyses and comparisons to these neural network models, they theoretically and numerically proved that the proposed NNN models possess the better convergence performance than that of the GNN and ZNN models for the real-time matrix Moore-Penrose inverse solving [8]. As a continuation of their works, in this paper, we further extend the NNN models to the more general cases named as Improved Recurrent Neural Network (IMRNN) models, which can outperform the existing GNN, ZNN and NNN models under certain conditions. Similar to the other papers, here we also highlight the main contributions and novelties of this paper as follows.
(1) Two IMRNN models can be utilized to efficiently solve the Moore-Penrose inverse of an real-time full-rank matrix, and they can globally converge to the exact Moore-Penrose inverse with an arbitrary initial state and some special monotonically-increasing odd activation functions.
(2) The convergence rate of the proposed IMRNN models can be much faster than GNN, ZNN and NNN models through an appropriate choice of the parameters and k involved in the models.
(3) The influence of the involved parameters on the convergence rate of IMRNN models is theoretically analyzed in detail and some numerical experiments are performed to further substantiated our theoretical results.
(4) The application of our IMRNN model to the inverse kinematic control problem of redundant robot manipulator also indicates the superiority to the known GNN, ZNN and NNN models.
The rest of this paper is organized as follows. In Section 2, we recall some basic notions and useful results on Moore-Penrose inverse. In Section 3, the existed GNN, ZNN and NNN models for computing left or right Moore-Penrose inverse of full-rank matrix are described briefly. Two improved recurrent neural network (i.e., IMRNN-L and IMRNN-R) models are proposed in Section 4, including the theoretical analyses of their superior convergence performance. In Section 5, some numerical experiments are performed to validate the superiority of IMRNN models. Finally, as a practical application and the testing of effects, our IMRNN models are applied to the kinematic control of redundant robot manipulator in Section 6.
Section snippets
Moore-Penrose inverse: notions, formulae and necessary properties
In this section, we briefly state the definition of Moore-Penrose inverse of a matrix and some formulae of a full-rank matrix. Some necessary properties related to the Moore-Penrose inverse are also mentioned. Definition 2.1 [27], [11] Let . Then the existed and unique matrix satisfying the following four matrix equationsis called the Moore-Penrose inverse of A and denoted by , where the superscript T denotes the transpose of a matrix or vector.
Specially, if a given
Some existed neural network models
To describe our improved recurrent neural networks, in this section we first state some existed neural networks for solving the Moore-Penrose inverse of a full-rank matrix.
Improved recurrent neural network (IMRNN) models
Based on the NNN models, in this section we proposed two improved recurrent neural networks i.e., IMRNN-L and IMRNN-R models for solving the Moore-Penrose inverses of the full-rank matrices.
The IMRNN-L model for the left Moore-Penrose inverse could be formulated aswhile IMRNN-R model for the right Moore-Penrose inverse could be written aswhere the designed parameters satisfy and that is integer,
Illustrative examples
In this section, we perform some numerical experiments on our proposed IMRNN models (16) and (17). However, the dynamic equations of the above neural network models are all described in matrix forms, which cannot be directly numerically simulated. Thus, we firstly transform the matrix-form differential equations into the vector-form differential equations via the Kronecker product and vectorization techniques [30], [31]. For eaxmple, the IMRNN-L model (16) can be transformed to the vector-form
Application to redundant-manipulator kinematic control
In this section, we apply the proposed IMRNN models to the minimum norm (MVN) scheme in the redundant-manipulator kinematic control with a three-link planar robot manipulator. The kinematic structure of such a robot manipulator is shown in Fig. 7.
Consider a three-link redundant robot manipulator of which the relationship between the end-effector position vector and the joint angle vector is embodied through the following forward kinematic equation
where the
Conclusions
In this paper we extend the known NNN models to the more general ones by introducing two parameters and k which can greatly improve the convergence performance of the NNN models by increasing the parameter values of and k when computing the Moore-Penrose inverse of a full-rank matrix. Our proposed new models are named as the improved recurrent neural network (IMRNN) models, i.e., the IMRNN-L and IMRNN-R models for solving the left and right Moore-Penrose inverses, respectively. We analyze
CRediT authorship contribution statement
Wenqi Wu: Conceptualization, Methodology, Software, Writing - original draft. Bing Zheng: Supervision, Project administration, Resources, Validation, Formal analysis.
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.
Acknowledgements
The authors thank the anonymous referees for their constructive suggestions and comments which greatly improve the presentation of this paper. This work is supported by the National Natural Science Foundation of China (Grant No. 11571004).
Wenqi Wu received her B.S. degree from School of Mathematics and Statistics, Heze University, Heze, China, in 2018. She is now pursuing his M.S. degree at School of Mathematics and Statistics, Lanzhou University, Lanzhou, China. Her research interests include neural networks, generalized inverse of matrix, and numerical linear algebra.
References (37)
- et al.
Iterative method for computing the moore-penrose inverse based on penrose equations
J. Comput. Appl. Math.
(2011) - et al.
Performance evaluation of feed-forward neural network with soft computing techniques for hand written english alphabets
Appl. Soft Comput.
(2011) - et al.
A power-type varying gain discrete-time recurrent neural network for solving time-varying linear system
Neurocomputing
(2020) - et al.
A recurrent neural network for computing pseudoinverse matrices
Math. Comput. Model.
(1994) Recurrent neural networks for computing weighted moore-penrose inverse 1
Appl. Math. Comput.
(2000)- et al.
A new noise-tolerant and predefined-time ZNN model for time-dependent matrix inversion
Neural Networks
(2019) - et al.
A parallel computing method based on zeroing neural networks for time-varying complex-valued matrix moore-penrose inversion
Inf. Sci.
(2020) - et al.
Nonconvex projection activated zeroing neurodynamic models for time-varying matrix pseudoinversion with accelerated finite-time convergence
Appl. Soft Comput.
(2018) - et al.
Li-function activated ZNN with finite-time convergence applied to redundant-manipulator kinematic control via time-varying Jacobian matrix pseudoinversion
Appl. Soft Comput.
(2014) - et al.
Effective neural remedy for drift phenomenon of planar three-link robot arm using quadratic performance index
Electron. Lett.
(2008)
A subspace approach to blind space-time signal processing for wireless communications
IEEE Trans. Signal Process.
Singular value decomposition and the moore-penrose inverse of bordered matrices
Siam J. Appl. Math.
Variants of the greville formula with applications to exact recursive least squares
Siam J. Matrix Anal. Appl.
Improved recurrent neural networks for online solution of moorepenrose inverse applied to redundant manipulator kinematic control
Asian J. Control
New discrete-solution model for solving future different-level linear inequality and equality with robot manipulator control
IEEE Trans. Ind. Inf.
Zhang neural network solving for time-varying full-rank matrix moore-penrose inverse
Computing
Cited by (20)
Zeroing neural network approaches for computing time-varying minimal rank outer inverse
2024, Applied Mathematics and ComputationDiscrete gradient-zeroing neural dynamics for future Moore–Penrose inverse with application to tracking control of manipulator
2023, Expert Systems with ApplicationsDirect derivation scheme of DT-RNN algorithm for discrete time-variant matrix pseudo-inversion with application to robotic manipulator
2023, Applied Soft ComputingCitation Excerpt :However, generally speaking, continuous time-variant algorithms may be more difficult to implement in industry than discrete time-variant algorithms, such as on the digital circuits and digital computers. Many researchers proposed improved RNN algorithms for solving discrete time-variant matrix pseudo-inversion after Zhang et al. introduced the algorithm [27], such as Wei et al. [28], Guo et al. [29], Liao et al. [30] and Petković et al. [31]. Besides, the study and processing of some mathematical problems can also be realized by RNNs [32–37].
A novel hybrid Zhang neural network model for time-varying matrix inversion
2022, Engineering Science and Technology, an International JournalCitation Excerpt :Xiao et al. [22] proposed a complex valued RNN model for time-varying complex matrix inversion which converges faster than CVZNN [19]. An improved RNN model for solving Moore-Penrose inverse of real-time full-rank matrix was investigated in [23]. The remaining part of the paper is arranged as follows: The problem statement is described in Section 2.
Wenqi Wu received her B.S. degree from School of Mathematics and Statistics, Heze University, Heze, China, in 2018. She is now pursuing his M.S. degree at School of Mathematics and Statistics, Lanzhou University, Lanzhou, China. Her research interests include neural networks, generalized inverse of matrix, and numerical linear algebra.
Bing Zheng was born in 1963. He received the M.S. degree in fundamental Mathematics from the School of Mathematical Sciences, Anhui University, China, in 1989, and Ph.D. degree in computational mathematics from the College of Sciences, Shanghai University, China, in 2003. He is currently a full time professor of the School of Mathematics and Statistics at Lanzhou University, China. His research interests include numerical linear algebra and its applications, multi-linear algebra, optimization theory, total least squares problem, perturbation analysis of the matrix and so on. He has published more than 80 papers in scientific journals.