THESIS PROPOSAL

HARDWARE EVOLUTION OF A ROBOT USING GENETIC

Submitted By:

Natasha Waris

Registration No.:

2015-Ms-Mc-06

Status:

Full time

Date of Registration:

26-01-15

Supervisor:

X-Y-Z

_____________________________________ ___________
Mechatronics & Control Engineering Department
University Of Engineering & Technology, Lahore

Hardware Evolution of a Robot Using Genetic Algorithm in Matlab

1 Introduction
In the last decades, natural computing methods which take problem solving principles
from nature have gained popularity. Among others, natural computing includes evolutionary
computing. Evolutionary algorithms often perform well approximating solutions to all types
of problems because they ideally do not make any assumption about the underlying fitness
landscape; this generality is shown by successes in fields as diverse as engineering, art,
biology, economics, marketing, genetics, operations, research, robotics, social sciences,
physics, politics and chemistry. Evolutionary computing covers population based, stochastic
search algorithms inspired by principles from evolution theory. An evolutionary algorithm
(EA) tries to solve a problem by keeping a set (population) of candidate solutions
(individuals) in parallel and improving the quality (fitness) of the individuals over a number
of iterations (generations). To form a new generation, genetically-inspired operators such as
crossover and mutation are applied to the individuals. A fitness-based selection process steers
the population towards better candidates.
Evolutionary algorithm is an umbrella term used to describe computer-based problem
solving systems which use computational models of evolutionary processes as key elements in
their design and implementation. EAs are computer programs that attempt to solve complex
problems by mimicking the processes of Darwinian evolution. In an EA a number of artificial
creatures search over the space of the problem. They compete continually with each other to
discover optimal areas of the search space. It is hoped that over time the most successful of these
creatures will evolve to discover the optimal solution. The artificial creatures in EAs, known as
individuals, are typically represented by fixed length strings or vectors. Each individual encodes a
single possible solution to the problem under consideration. For example, in order to construct an
EA to search the conformation space of a molecule, each angle of rotation around a flexible bond
could be encoded as a real number. Concatenating these numbers gives a string which can be
used within an EA. Thus, each individual would encode a specific set of torsion angles. EAs
manipulate pools or populations of individuals. The EA is started with an initial population of
size comprising random individuals (that is, each value in every string is set using a random
number generator). Every individual is then assigned a fitness value. To generate a fitness score
the individual is decoded to produce a possible solution to the problem.
The value of this solution is then calculated using the fitness function. Population
members with high fitness scores therefore represent better solutions to the problem than
individuals with lower fitness scores. Following this initial phase the main iterative cycle of the

algorithm begins. Using mutation (perturbation) and recombination operators, the individuals in
the current population produce children. The children are assigned fitness scores. A new
population of individuals is then formed from the individuals in the current population and the
children. This new population becomes the current population and the iterative cycle is repeated.
At some point in the cycle evolutionary pressure is applied. That is, the Darwinian strategy of the
survival of the fittest is employed and individuals compete against each other. This is achieved by
selection based on fitness scores, with fitter individuals more likely to be selected. The selection
is applied either when choosing individuals to parent children or when choosing individuals to
form a new population.

2 Literature Review
Until last few decades not a lot of work was done on evolutionary algorithms but in
these days researchers are taking interest in this area providing some bases to start research
on this topic. Proposed research will describe evolutionary algorithms and will focus on one
of the evolutionary algorithms called genetic algorithm (GA).
This paper [1] gives the basic idea on how to evolve different embedded components
using different techniques with GA and how to make a comparison between them. Paul
Kaufmann and Marco Platzner described that evolutionary hardware design reveals the
potential to provide autonomous systems with self-adaptation properties. An architectural
concept for an intrinsically evolvable multi-objective (e.g., functional quality, area and speed)
hardware that adapts to slow changes in the environment by simulated evolution, and to rapid
changes in available resources by switching to pre-evolved alternative circuits. Long-term
goal of this paper is the development of autonomous embedded systems that implement
hardware functions (circuits) characterized by their functional quality and resource demand.
Three concepts to achieve a flexible adaptation are intrinsic evolutionary search process, preevolved alternative which meets the new resource constraints and a reconfigurable system on
chip platform is the technology allowing for the replacement of circuits during runtime.
Cartesian Genetic Programming (CGP) introduced by Miller and Thomson is used in this
work. CGP is a structural hardware model where a circuit is formed by combinational logic
blocks arranged in a two-dimensional array and an interconnect (wires) between the blocks.
In the main part of the paper, evolutionary circuit design was treated as a multi-objective
optimization problem and compares two multi-objective optimizers (SPEA2 and TSPEA2)
with a reference genetic algorithm. In the experiments, the best results were achieved with
TSPEA2, an optimizer that prefers a single objective while trying to maintain diversity.
A comparison between Genetic Algorithm and other evolutionary techniques shows
the advantages and importance of GA. Tetsuya Higuchi [2] made a comparison between
neural networks and evolvable hardware (EHW). EHW is a hardware device which can adapt
its own hardware structure to the environment to give the best performance. Subsumption
architecture is a new approach to building behavior-based robots. In such a robot, a goal is
attained through interactions among simple modules. Adding a new module to a robot leads
to improving the performance of a robot. However, in that case, a mechanism is required to

arbitrate the interactions between the new module and existing modules. However, as the
number of modules increases, the arbitration becomes more difficult. It is almost impossible
for programmers to specify all the occasions where arbitration is necessary when the robot
contains many modules. This is one of problems associated with the subsumption
architecture. EHW is one of the promising candidates to solve this.
Adaptation in real-time is feasible due to a speed-up by many orders of magnitude. The
system will be flexible and fault-tolerant since EHW can change its own structure in the case
of environmental change or hardware error. EHW has the pattern recognition capability like
neural network. In addition to it, EHW can attain real-time performance which could not be
obtained by neural networks.
The aim of this paper [3] is to extend evolutionary techniques into the physical world
by evolving diverse robots that can be fabricated automatically that are functional in the real
world, diverse in architecture, and producible in their entirety at low cost and large quantities.
The author then describes a set of preliminary experiments examining the ability to achieve
fully automated design and transfer into reality, both with as little human intervention as
possible. Jordan B. Pollack, in his paper proposed an approach that is based on the use of
only elementary constituents in both the design and fabrication process. As building blocks
become more elementary, external knowledge associated with them is minimized, and at the
same time architectural flexibility is maximized. In a set of experiments bars are used as
building blocks of structure, neurons as building blocks of control, and additive fabrication as
a production process. Bars connected with free joints can form trusses that represent arbitrary
rigid, flexible and articulated structures as well as multiple detached structures, with revolute,
linear and planar joints at various levels of hierarchy. With a population of 200-1000
machines that were initially comprised of zero bars and neurons, evolution was conducted in
simulation. The fitness of a machine was determined by its locomotion ability and then
automatically replicated into reality using 3D solid printing. The resulting structures
contained complex joints that would be difficult to design or manufacture using traditional
methods. In spite of the relatively simple task and environment (locomotion over an infinite
horizontal plane), surprisingly different and elaborate solutions were evolved. Machines
typically contained around 20 building blocks, sometimes with significant redundancy. The
ability to move artificial evolution from simulation to the real world is not a mere curiosity;
rather, some claim that if indeed artificial systems are to ultimately interact and integrate with
reality, they must learn, evolve and be studied in it.
The plus point of this method is that it is very useful in evolving complex structures with
multiple bars and joints that are actually working but this method runs a lot of simulations so
its time consuming and when simulated results were converted into reality there were some
short comings like Newtonian friction was ignored so the robot was slipping in reality.
This thesis [4] puts forward new techniques for evolving control systems for real
robots using easy-to-build, fast-running simulations. It begins with a tutorial in state-of-theart Evolutionary Robotics and discusses the best types of neural network, encoding scheme,
genetic algorithm and genetic operators to use - and how to use them. Several novel types are
introduced, and their relative merits over previous approaches are discussed. A new

methodology is proposed for building minimal simulations within which controllers that meet
these conditions will successfully transfer into reality.
Building a minimal simulation for Evolutionary Robotics proceeded in three stages:
1. The inevitable differences between simulation and reality were examined, and the
reasons why some controllers can overcome these differences to successfully transfer
into reality while others cannot were identified. This lead to the development of two
conditions that controllers must fulfill if they are to cross from simulation into reality.
2. In order for it to be possible that controllers can fulfill these conditions, a simulation
must model certain real-world features and processes, and these were identified. A
simulation that models no more than the minimum necessary was labelled a minimal
simulation.
3. By no means will every controller evolved in a minimal simulation fulfill the
conditions for successful transfer, and techniques were proposed for using the
evolutionary process itself to force evolving controllers to meet them. Controllers are
not only evolved to perform a specific behavior within a minimal simulation,
therefore, they are also evolved to fulfill the conditions for successful transfer into
reality.
If controllers are to perform the same behavior in reality as they do in simulation, a new
methodology is proposed for building minimal simulations within which controllers that meet
these conditions will successfully transfer into reality. Techniques are then put forward for
forcing controllers that evolve to be reliably fit within such minimal simulations to meet these
conditions. Four sets of experiments are reported, all involving minimal simulations.
I.

II.
III.

Controllers were evolved for a small mobile robot that could solve a T-maze in
response to a light cue, target recognition and approach behaviors were evolved
for a visually guided mobile robot,
Walking and obstacle-avoiding behaviors were evolved for an eight-legged robot
Motion-tracking behaviors were evolved for a simple panning camera head

In all cases, the evolution of complex robot behaviors that would have taken months to
evolve if fitness evaluations had been performed in reality was performed in a matter of
hours, and controllers that evolved to be reliably fit in simulation displayed extremely robust
behavior when downloaded into reality. This is an extremely complex behavior to evolve by
todays standards of what can and cannot be evolved, and yet the necessary minimal
simulation remains simple and fast. The point is that whether a minimal simulation is easy to
construct and runs fast depends not on the complexity of the behavior we want to evolve
using it, nor on the complexity of the robot that it simulates, but only on the complexity of the
base set of environmental features necessary to underlay the behavior for that robot.

This paper [5] introduces evolvable hardware (EHW) and surveys the work on EHW
and further describes its possible applications. One of the goals of the author is to explain
how hardware evolution can be achieved by using the local behavior of the circuit

components instead of global behavior and then global behavior emerges from local behavior
giving the system unlimited scalability. To achieve this Genetic Algorithm (GA) is used. In
which the members of a population are initially randomly generated which are called
chromosomes or genotype. The members are evaluated and sorted on fitness according to a
given fitness criteria. The parameters of fittest chromosomes are exchanged to generate the
next generation which is preferably fitter. Mutation may also occur and introduce changes
that may not be present in parents. GA is used for optimization problems and for adaptive
systems for robot control.
Large applications proposed for EHW are -Autonomous robots: vacuum cleaner, surveillance
robots, automatic de-mining robots, maintenance robots, -Pattern recognition, -Signal
processing.
The Evolvable system may adopt several behavior characteristics of natural living systems
like emergent behavior, adaption to the environment, self-replication and self-reparation.
EHW can solve difficult problems in these applications better than other methods.
Evolvable Hardware can solve difficult problems better than other methods like artificial
neural networks. Autonomous robots require very complex systems to be functional and
evolving such a system instead of designing and programming it, the complexity level can be
highly increased. So to use such a system in real time evolvable hardware is required. It has
several advantages of fault tolerance, abstraction, scalability, evolution without explicit
function using some objective fitness function.

3 Problem Statement:
There are several good reasons for evolving the hardware of a robot. The physics of
interaction between the robot and the environment are often difficult to anticipate and to
include in software models and for this reason evolutionary development of the observed
behavior and of the control mechanism that supports it is carried on. Evolutionary robots
could only adapt the software running within their processor, but could not change the
physical properties of their circuits, sensors, motors or even the shape of their body.
This dissertation will seek to explain the development of a robot which can turn on
and off its sensory module in accordance to the changes in its dynamic environment. A
dynamic environment consists of changing surroundings in which the robot navigates. So
unlike the static case, the robot must adapt to new situations and overcome possibly
unpredictable obstacles. So the aim will be to develop a global control of robot navigation
and the evaluate it to perform genetic algorithm and achieve hardware evolution of the robot
in terms of turning the sensors ON & OFF.

4 Proposed Solution:
Nowadays Robot Simulation Softwares are of prime importance, and they may be
designed for null cost. Thus, the only expense in designing of a robot is in designing
mechanical structure of the robot. Thus for this research the robot will be simulated in Matlab

and after creating the genetic algorithm for the robot it will be tested and the results will be
compared with the help of simulations and the graphs generated by the Matlab. Genetic
Algorithm toolbox in Matlab will also be used for creating genetic algorithm (GA). Since
this algorithm is iterative in nature so Matlab simulations are the best.
The research will be conducted in following sequence:
Study of existing work done on evolutionary algorithms, genetic algorithm and
hardware evolution
Simulation of the robot using Matlab
Development of genetic algorithm using GA toolbox in Matlab
Testing the robot in simulation with GA
Analyzing and comparing the results of my simulation
Compiling the results of simulation effectively and complete the research work.

5 Intended Results
The aim of this research is to develop a robot with evolution properties for unstructured
environment. This robot can be a multi-purpose robot. From industrial point of view
constructing different robots for different tasks can take a lot time and effort but this robot
can be used as a multitasking robot so it will save a lot of time on an assembly line.

6 Time Table
Considering the scope of work, at least 3-4 weeks are required for this research. This will be
arranged according to the following.
1-2
Literature Review
Data Collection and Analysis
Programming & Simulation
Verification of results
Thesis Report and presentation

2-4

Time Schedule Weeks

10-12

References:

[1] Kaufmann, Paul, and Marco Platzner. "Toward self-adaptive embedded systems: multiobjective hardware evolution." Architecture of Computing Systems-ARCS 2007. Springer
Berlin Heidelberg, 2007. 199-208.
[2] Higuchi, Tetsuya, et al. "Evolvable hardware with genetic learning." Circuits and
Systems, 1996. ISCAS'96., Connecting the World., 1996 IEEE International Symposium on.
Vol. 4. IEEE, 1996.
[3] Pollack, Jordan B., and Hod Lipson. "The GOLEM project: Evolving hardware bodies
and brains." Evolvable Hardware, 2000. Proceedings. The Second NASA/DoD Workshop on.
IEEE, 2000.
[4] Jakobi, Nick. Minimal simulations for evolutionary robotics. Diss. University of Sussex,
1998.
[5] Torresen, Jim. "Evolvable hardwarethe coming hardware design method."Neuro-Fuzzy
Techniques for Intelligent Information Systems (1998): 435-449.