Abstract—The solving strategy of artificial intelligence (AI) is
adopted with bottom-up design to solve its hard problems. To
tackle end-to-end AI-hard problems, a highly self-adaptive
control system-on-chip has been developed to self-learn its
internal and external resources with the aid of sets of sensors
and actuators. Inspired by biological cell learning theory,
different approaches of modelling techniques have been derived
together with machine learning methods to the embedded
control systems so as to perform different tasks. Some
experimental results have shown the developments.
I. INTRODUCTION
ith its strong inter-disciplinary characteristics, machine
learning is the study of computer algorithms that
improve automatically through experience [9]. Many
different machine learning (ML) techniques have been
developed recently. Decision trees are used in knowledge
engineering and data mining applications, expert systems can
help tackle specific problems that require comprehensive
knowledge. Fuzzy sets can calibrate vagueness, as required
for natural human language processing.
Artificial neural networks (ANNs) are inspired from
interconnections of brains cells. They are ‘processing
memories’ where the knowledge is distributed over tiny units
that operate in parallel. Algorithms inspired from genetic
processes can be applied to solve any problem in which the
solutions can be encoded. GA (Genetic Algorithm)
approaches can be taken when a “good enough” solution is
required quickly, but without the need to understand the
rationale of this particular solution(s). Indeed, to guide the
evolution of a GA, we preliminary need to define the fitness
function or landscape function that corresponds to the given
problem we wish to tackle.
The main goal of such algorithms is to increase the
machine knowledge, while other goals are to allow the
machine to achieve tasks based on the recorded knowledge
(i.e. the experience) and on the currently perceived
Y. Zhou is with Shenzhen Institutes of Advanced Technology, Chinese
Academy of Sciences, and also with Shenzhen Key Laboratory of Electric
Vehicle Powertrain Platform and Safety Technology, Shenzhen, China,
518055 (phone: 86-755-86392152; fax: 86-755-86392194; e-mail:
ym.zhou@ siat.ac.cn).
L.Krundel, D. Mulvaney and V. Chouliaras are with Department of
Electronic and Electrical Engineering, Loughborough University,
Loughborough, UK, LE11 3TU (e-mail: l.a.krundel, d.j.mulvaney,
v.a.chouliaras@lboro.ac.uk).
G. Xu is with Shenzhen Insitutes of Advanced Technology, Chinese
Academy of Sciences, and also with The Chinese University of Hong Kong,
Hong Kong (email: gq.xu@siat.ac.cn).
G. Fu is with School of Mechanical, Electronic and Control Engineering,
Beijing Jiaotong University, Beijing, China (e-mail: gqfu@bjtu.edu.cn).
surrounding information, and perhaps ultimately to allow the
machine to think [4,5] in order to permit the solution of
complex problems that would not normally be possible to
tackle.
Conventional Cellular Automata (CA) is a natural way of
studying the evolution of large physical systems and
constitute a general paradigm for parallel computation [14].
They are clocked: whole 2D CA landscape updated at each
clock tick. They can be of nth-order if they are made of n
interacting 2D cell layers. Each cell can usually have two
possible states: e.g., alive or dead. They are governed by local
(action-at-a-distance forbidden) and uniform (single rules’
recipe) system’s rules. Interestingly, one can define a
meta-rule, which consists of several sets of rules and a
strategy to dictate which rule’s set applies and when in the
CA progress. This can bring up sophisticated behaviours.
Cellular automata have properties that make them suitable
for bottom-up design [14], including the generation of
emergent NN structures. In highly distributed, massively
parallel reconfigurable machines (such as CAs), spontaneous
organization and emergence of intelligent behaviours can
occur [11].
As systems grow in complexity, there is a huge need for
features such as self-management of system resources.
Unfortunately, conventional computing systems have to
incorporate this feature as an additional software application
that is cumbersome or impossible to maintain. However,
redundant cellular systems are inherently scalable and
self-management is straight forward to design and implement
at elaboration time, and to maintain at simulation/run time.
CAs also lend themselves to self-reorganization and can be
easily restructured at run-time. Finally, in CAs the processing
units are distributed, which ensures better fault-tolerance and
handling of indeterminism. This paper discusses the CA
learning rules and preliminary results.
The reminder of the paper is organized as follows. In
section II, background of the learning methods in neural
networks is introduced. Methods of self-configuration are
discussed in section III. The theory and learning rules of
cellular automata technique and related proposed projects
based on CA rules are explained in section IV. Preliminary
results are given in section V.
II. BACKGROUND
A. Review Stage
Autonomous learning design in System-on-chip
Yimin Zhou, Member, IEEE , Ludovic Krundel, Member, IEEE , David Mulvaney, Vassilios Chouliaras,
Guoqing Xu, Member, IEEE , Guoqiang Fu
W
978-1-4799-2744-9/13/$31.00 ©2013 IEEE
Proceeding of the IEEE
International Conference on Robotics and Biomimetics (ROBIO)
Shenzhen, China, December 2013
1054