Complex-valued Kernel Incremental Metalearning
Algorithms
Yan Ye and Chunuang Li
Department of Information Science and Electronic Engineering,
Zhejiang University,
Hangzhou 310027, People’s Republic of China
Email: cgli@zju.edu.cn
Abstract—Metalearning algorithm learns the base learning
algorithm, thus to improve the performance of the learning
system. Usually, metalearning algorithm exhibits faster conver-
gence rate and lower mean-square error (MSE) than the original
base learning algorithm. The Kernel method is a powerful
tool for extending an algorithm from linear to nonlinear case.
In a previous work, we have presented a kernel incremental
metalearning algorithm (KIMEL). In recent years, complex-
valued signal processing algorithms are gaining popularity due
to their broad applicability. In this paper, we present complex-
valued KIMEL (CKMIEL), which has two versions. One is based
on the complexification of the real RKHS, named CKIMEL1,
while the other uses a pure complex kernel, named CKIMEL2.
To demonstrate the effectiveness and advantage of the proposed
algorithms, we apply them to nonlinear channel identification.
Experimental results show that the CKIMEL algorithms have fast
convergence rate and low convergence MSE, and the CKIMEL1
algorithm is superior to the competing algorithms.
Index Terms—Kernel, Complex, Metalearning, Adaptive Ker-
nel Learning, Channel Identification
I. INTRODUCTION
Matelearning means learning of the base learning system,
hence to improve the performance of the base learning system.
The delta-Bar-delta (DBD) [1] and the incremental delta-bar-
delta (IDBD) [2] are well-known metalearning algorithms,
which consist a weight update and a learning rate update
both based on the delta rule. They are developed for the
linear learning systems. If the mapping between output and
input is highly nonlinear, their performance is poor. In [3],
we developed a Kernel Incremental MEtaLearing (KIMEL)
algorithm to overcome the linearity limitation.
However, most of the kernel-based algorithms [4], including
the KIMEL [3], were designed to solve real-valued prob-
lems. Complex-valued signal processing is gaining popular-
ity due to their broad applicability. Recently, some kernel-
based techniques were proposed for processing complex-
valued signals. In [5], the authors presented two frameworks
based on complex reproducing kernel Hilbert space (RKHS)
to solve the complex-valued kernel learning problem. The
main contributions of the paper include: a) the extension
of Wirtinger’s Calculus in complex RKHS as an efficient
means for computing the gradients, which are involved in
the derivation of complex adaptive learning algorithms; b) the
development of a framework that allows real-valued kernel
algorithms to process complex signals by taking advantage of
the complexification of real RKHS; c) using pure complex
kernels for kernel-based adaptive processing of complex data.
By using these tools, they proposed two complex-valued forms
of the KLMS algorithm, named CKLMS1 and CKLMS2,
respectively.
The KIMEL algorithm proposed in [3] is an adaptive
kernel learning algorithm with a variable step-size, which has
better performance than the KLMS. In this paper, using the
frameworks proposed in [5], we extend the KIMEL algorithm
to the complex RKHS and present two complex versions of the
KIMEL, named CKIMEL1 and CKIMEL2, respectively. The
former takes advantage of the complexification of real RKHS,
while the later uses the pure complex reproducing kernels.
Note that the CKIMEL algorithm is not simply a CKLMS
algorithm with a variable step size. Since the learning-rate
update rule depends on the inputs, it is also updated in complex
RKHS.
The rest of this paper is organized as follows. In Section
II, we briefly introduce the KIMEL algorithm, the complex
RKHS and Wirtinger’s Calculus, to make the paper self-
contained. In Section III, the CKIMEL algorithms are for-
mulated. In Section IV, an application on nonlinear channel
identification is presented to illustrate the effectiveness and
advantage of the proposed algorithms. Finally, conclusions are
drawn in Section V.
II. PRELIMINARY KNOWLEDGE
In this paper, the proposed new algorithms are based on
the KIMEL algorithm and the complex kernel method, thus
in this section we briefly introduce the corresponding knowl-
edge on complex RKHS, Wirtinger’s Calculus and KIMEL.
Throughout the paper, we denote the set of all integers, real
and complex numbers by N, R and C, respectively.
A. Complex Reproducing Kernel Hilbert Space
A RKHS is defined as a Hilbert space H over a field F
for which there exists a function κ: X ×X → F [6], [7]. The
function κ is positive definite, which has the following two im-
portant properties: a) For every x ∈ X, κ(·, x) belongs to H;
b) κ has reproducing property, i.e., f(x) = ⟨f, κ(·, x)⟩
H
, for
all f ∈ H, x ∈ X, in particular κ(x, y) = ⟨κ(·, y), κ(·, x)⟩
H
.
The map Φ: X ∈ H : Φ(x) = κ(·, x) is called the feature
978-1-4799-2186-7/14/$31.00 ©2014 IEEE