Привіт Гість ( Вхід | Реєстрація )

 
Reply to this topicStart new topic
> Alife@home, научные эксперименты касающиеся нейросетей
nikelong
Aug 25 2008, 11:10
Пост #1


Тера ранчер
**********

Група: Trusted Members
Повідомлень: 12 443
З нами з: 19-March 05
Користувач №: 92
Стать: Чол





Проект "Alife@Home"

--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
ТОП-20 участников:
----------------------------------------------------------------------------------------------------------
Дата основания команды - Дата Капитан - Капитан
----------------------------------------------------------------------------------------------------------
Полезная информация:
Для идентификации пользователя в BOINC могут служить 2 вещи:
1) пара e-mail/пароль
2) межпроектный идентификационный ID (Cross-project ID) - 32значное шестнадцатиричное число.

Если Вы пожелаете подключится ещё и к другому BOINC-проекту, то помните: чтобы не плодить новых аккаунтов при подключении к новому проекту или команде, нужно обязательно везде регистрироваться с одним и тем же e-mail/паролем либо CPID. если при регистрации в проекте указать другие e-mail или пароль, BOINC создаст новый аккаунт с тем же именем!
----------------------------------------------------------------------------------------------------------


О проекте:
Проект стартовал 12.10.2004
Short for 'Artificial Life at Home' is an effort to conduct scientific experiments regarding neural networks and evolution on the computers of volunteers. This way, unused computer power in many homes, and therefore wasted electricity, could be used for something meaningful. At the same time, the researcher can do experiments that would otherwise be unfeasable due to the amount of required computer equipment. If you are interested, you can read some general introduction to the field, you can read about the research done at ALife@Home, you can read something about the technical details regarding distributed computing, or you can go straight ahead and open an account to donate computer power to ALife@Home.

ALife@Home: Introduction
(Show/Hide)

ALife@Home is a project that allows to conduct scientific experiments regarding neural networks and evolution
on the computers of volunteers. This page is meant as a very low level introduction into (artificial) neural networks, (artificial)
evolution and the objectives in this kind of research. If you know what these terms mean, you can probably skip this page and
visit the research section, where the scientific background of the experiments running on ALife@Home is explained. If you are
interested in a particular experiment that is running at the momemt, please visit the research blog.
Neural Networks

The brains of animals are the most complex structures in the known universe. The 20th century shed lots of light on two levels
of undarstanding of the whole brain:

On the one hand, we know a lot about the location of different information pathways. We know what is processed where, and
which parts have to be intact for that function.

On the other hand, neurophysiologists can describe in great detail how single neurons and the connections between them work.
The level in between however, which describes the complex interactions in populations of neurons, is still a field of wild
speculations.
Evolution

As can be seen in offspring of animals and plants, children of a given parent are not all the same, but differ individually. Two
mechanisms are responsible for this: recombination of genes and mutation of genes.

If we assume some of the resulting genetic configurations to be more suitable for the production of offspring than others, these
configurations are represented in the next generationn with a higher frequency than before. As can be seen for example in the
different dog breeds, little changes can over long time spans accumulate to substantial differences.

If populations are separated for a long time, and accumulate so many differences that they are not able to successfully reproduce
anymore, a new species is created.
Artificial Neural Networks (ANNs)

Scientists agree that we could model a whole brain if we completely understood how it works and employed arbitrarily complex
neural models (e.g., if we modelled every single atom in a brain). To model every atom in a brain, however, is impossible,
unfeasible and not very interesting, since the process of building a machine that models every atom in a brain would require
much more knowledge than understanding the brain itself. The hope is, that simpler models can be found that are sufficient to
model some parts of a brain on a functional level. It is however unclear, how complex the neural models have to be to reach
that goal.

The most commonly used artificial neuron (neuroid) is a simple summation unit that adds up weighted inputs of other neuroids.
With this kind of ANNs, already very complex functions can be achieved. They are used in technical systems to recognize
fingerprints, faces, speech and handwritings, to control robots and to kategorize things. Nevertheless, these are very specialized
instances, and they show no particular similarity to biological brains.

Ofcourse, there exist also models that integrate arbitrary amounts of our knowledge on the molecular level. In consequence,
they take much longer to develop and run much slower, a disadvantage that was so far not overcome by some great new
functionality.

The optimal solution lies somewhere in between: a model that integrates only the esential features of a biological neuron without
simulating parts that are either unnecessary, redundant, or only due to biochemical pecularities of the biologic world.
Approaches are: (1) Building more and more complex models of the single neuron by integrating as many as possible findings
of the biological research. Only if we can model one neuron more or less exactly, we can be sure that we already know the
mechanisms we are looking for. (2) Trying to improve existing, complex network models with single biological findings. If we can
improve the functionality, we have probably found an esential mechanism by accident.
Artificial Evolution

Evolution is a ubiquitous process. If a car manufacturer tentatively produces variations of a given car, and then selects the
one using the least fuel, the car "evolves". As soon as this process is automated, we speak of "artificial evolution". The same
can be applied to programs or Artificial Neural Networks (ANNs): Variations of a program or an ANN are produced, tested for
their performance on a given task, and afterwards the best solutions are selected for the next generation. The only difference
to biological evolution is that WE design the environment where solutions reproduce, whereas biological evolution happens in a
predifened environment called "the universe". Thus, evolution can easily be transferred to non-biological systems. This way, like
in biology, we can find solutions for problems without understanding the problem, but much faster than by just putting out
randomized solutions.

If we want to mimic biological evolution, however, we have a huge disadvantage: we would like to evolve something in a
reasonable time on a reasonable number of PCs. In contrast, biological organisms evolve in millions of years. Furthermore, as in
the reign of of procaryots (which are, compared to most computer models, impressively complex), every year
1,740,000,000,000,000,000,000,000,000,000 individuals are tested for their fitness (source).
Distributed computing with BOINC!

ALife@Home utilizes a framework called "BOINC". BOINC is an acronym for Berkeley Open Infrastructure for Network Computing
. The idea dates back to a project called SETI@Home, which is short for "Search for ExtraTerrestrial Intelligence at Home".
Researchers of this field were the first to discover the enormous potential of using otherwise unused computer power on privatly
owned PCs all over the world. After first being laughed at, the project quickly became the fastest computer on earth, much
faster than supercomputers worth millions of dollars.

BOINC is an effort to generalize this model, such that more research projects have the chance to compete with each other
for private computer power. If you contribute to one project, you can also attach to other projects and divide your resources
between them. In other words, these projects try to convince you to allow them to borrow your computer when you aren't
using it and to help them calculate some things.

There is probably lots you have to share: 99% of the time your computer is turned on, it is actually doing (almost) nothing:
waiting for you to move the mouse, waiting for the modem to deliver some data, waiting for the harddrive and so on. If you
contribute to some project, this project will send some application and some data to your BOINC client. Whenever your computer
is idle, the BOINC client starts one of the applications, which in turn begins to process its data. When the data is processed
completely, the results are sent back to the project, and new data is received.

As soon as you use your computer yourself, BOINC cuts down the amount of resources it is using, so in the optimal case
you should never realize it is there.
ALife@Home
ALife@Home is a project that allows to conduct scientific experiments regarding neural networks and evolution on computers
of volunteers. If you would like to contribute, you are asked to download and install a program called "BOINC! client". Then,
you can attach to different research projects, one of them being ALife@Home, and donate computer power to these projects.
Thats all to do on your side.

The client will work in the background and download an artificial world. This artificial world then gets populated by artificial
creatures with artificial genomes, that develop artificial neural networks in an artificial ontogenesis. They undergo an artificial
evolution, and, after a certain amount of time, statistical data on the evolutionary process together with the genome of the
fittest creature is sent back to ALife@Home.


Research Objectives
(Show/Hide)

This page describes the general objectives pursued at ALife@Home in a broad perspective. If you dont
understand anything at all, try the introduction. If you are interested in the brief description of a particular experiment that
is running at the momemt, please visit the research blog.
Evolving controllers for six legged walking

Background
If someone wants to control a robot, there are two principal approaches:

(S)he formulates kinematics and dynamics of the walking machine and explicitly calculates the motor commands necessary
to get the robot moving around. These programs are human readable and human understandable, but computationally
very expensive. This approach will further be called the classical approach.

Fig. 1: A two dimensional six legged walker

Fig. 1: A two dimensional six legged walker. The filled rectangle in the middle represents the body, the circles on both sides
represent the action ranges of the leg endpoints.

(S)he defines an artificial neural network, most likely involving some supervised learning methods, that afterwards controls
the robots behaviour. These neurocontrollers are very fast, but very difficult or even impossible to design and to analyze.
This approach will further be called the neural approach.

All commercial robots like asimo, aibo and qrio employ the classical approach, the neural approach is so far mainly restricted
to research projects. Since these neurocontrollers are so difficult to design, researchers are trying to employ self-organizing
principles, one of them being artificial evolution, to develop more complex controllers. The efforts undertaken so far can
mostly be grouped in two kategories:

(A) Punctiform entities are evolved on a two dimensional grid. This way, complex behaviours like orientation in a labyrinth,
landmark navigation or swarm behaviour could be evolved. These simulations usually do not contain any physics, which
is probably the reason for the difficulties encountered when implementing these controllers in hardware robots. Furthermore,
a 1:1 implementation of the evolved neurocontrollers is only possible for wheeled robots, since the control of legged machines
requires much more complex control structures. These restrictions further promote the importance of "Embodiment", i.e.,
the relevance of having a physical body.

(B) Embodied creatures are evolved in real world hardware, in more or less accurate dynamics simulations, or in combinations
of the latter. The evolved neurocontrollers can much more easily be transferred to robots. The proximity to real world physics
is paid with a tremendous increase in necessary computer power. Last but not least, it can be argued that intelligence as we
know it (or at least, as we can understand it) can only evolve in complex, embodied and situated agents with surplus degrees
of freedom.

The optimal strategy to evolve artificial life would thus be the co-evolution of a high number of complex embodied agents
with complex control structures in a complex environment under realistic physical constraints. This accounts for a heavy
trade-off in artificial life: both the requirements for a capable evolution AND the requirements for complex embodied agents
in complex environments require huge amounts of computational power. Thus, one can invest in the first OR the latter.

The neurocontrollers evolved according to approach B so far show a significant shortcoming: While nature evolved highly
reactive systems with intimate integration of body and brain, artificial evolution so far evolved mainly simple pattern generators
for motion control with sparse connections to higher levels. This was partly intented to find the simplest possible solution, but
partly it is due to the heavy trade-off between evolution and embodiment mentioned before.

Introduction
The experiments in this subproject are designed to build a bridge between the aforementioned extremes A and B. While the
agents still show essential properties of embodiment, the simulation is fast enough to sport a high number of individuals over
a high number of generations.

Technical Details

Central vs modular control

Fig. 2: Central and modular controllers of varying design are compared for their evolvability.
Agent design is that of a two dimensional six legged walker (see Fig. 1). This architecture was chosen for the vast amount
of biologogical data on the stick insect Carausius morosus, a three dimensional six legged walker. This data will be
compared to the data gathered with the two dimensional approach. Defined are the body, the idle position of each leg endpoint,
and the actionrange of a leg endpoint. Controlled are either

the AEP (anterior extreme position), the PEP (posterior extreme position) and the speed at which the leg endpoint is moving
in between or

the x-velocity and the y-velocity of the leg endpoint and a boolean that states if the leg is in swing or stance. The agents are
equipped with neural networks, which are to varying degree defined and evolved, respectively. They are either evolved to
walk as far as possible or to collect resources.
Objectives
Network structure

Different architectures of neural networks are compared for their evolvability. A general question asks for the advantages
and disadvantages of central controllers versus modular controllers (see Fig. 2). A more specific question asks for the optimal
number of intra-leg feedback connections and for the optimal number of inter-leg connections, respectively. These
architectures are compared in simple straight walk as well as in complex resource following behaviour.
Turning

The turning behaviour evolved while collecting resources is investigated. Which parameters are mainly influenced? How
does that compare with stick insect data (gathered by Wiebke Ebeling and Volker Duerr). How do modular and central
controllers differ in their curve walking behaviour.
Reversing

How do the two dimensional animats turn around at dead ends? How do they avoid obstacles? How does that compare to
stick insect data (gathered by Sebastian)
Gaps

Even in two dimensions, there can be gaps in the ground: areas, where leg points are not allowed to touch the ground.
If we define gaps in the ground that can be crossed: How do animats approach and cross a gap? How does that compare
to stick insect data (gathered by Bettina Blaesing).
Center of gravity

The center of gravity in the stick insect Carausius morosus is located between the hind legs. The center of gravity of Tarry,
a robotic platform based on stick insect morphology, is between its middle legs. How does a shift of the center of gravity
influence walking behaviour in the two dimensional agents? What conclusions can be drawn for the robot?
Operating range and asymmetric efficiency of individual legs

Insects as well as robots feature a predefined operating range and an asymmetric efficiency in respect to direction of
movement of the individual leg. How do these parameters for the two dimensional agents depend from other body constants
and from the task to accomplish?
Structural and funcional analysis of evolved controllers

In different evolutionary runs, different solutions ( = different local optima) are found. If we compare the controllers evolved
in different runs, can we find structural or functional similarities. Can we identify pattern generators or purely reactive
solutions? Does positive stance feedback, as proposed for the stick insect, evolve?
Evolution of targeting behaviour

If agents are evolved on a grid, i.e., a cyclic pattern of gaps and stable ground, do they evolve a target mechanism as
seen in the stick insect?
Homeostatic plasticity

Can neuroids endowed with homestatic plasticity (investigated for biological neurons by Turrigiano) balance noise during
ontogenesis?
Evolution of higher behaviour

Can we evolve higher behaviours like landmark navigation, obstacle avoidance and orientation in a labyrinth even with
six legged entities?




Ссылки по теме:Project Info

* Year of conclusion: 2005
* Number of teams: 37
* Number of participants: 251
* Running platforms: BOINC (Windows, Linux)

http://wiki.bc-team.org/index.php?title=Alife%40home/en


Це повідомлення відредагував nikelong: Aug 30 2010, 00:12
User is offlineProfile CardPM
Go to the top of the page
+Quote Post

Reply to this topicStart new topic
1 Користувачів переглядають дану тему (1 Гостей і 0 Прихованих Користувачів)
0 Користувачів:

 



- Lo-Fi Версія Поточний час: 29th March 2024 - 15:15

Invision Power Board v1.3.3 © 1996 IPS, Inc.