User:Elmruud

From Robin

(Difference between revisions)
Jump to: navigation, search
(Contact)
 
(3 intermediate revisions not shown)
Line 6: Line 6:
== Description of master project ==
== Description of master project ==
 +
Lifetime Learning - Machine learning in an evolutionary robotics context
 +
In evolutionary computing, there are two main models for the application of local search or machine
 +
learning during the lifetime of an individual: Lamarckian learning, where the genes of an individual is
 +
updated so that lifetime experiences gets passed on to new generations, and Baldwinian learning,
 +
where only the fitness of an individual is changed. Baldwinian learning is closest to the the way the
 +
evolutionary process in nature is understood today, but in evolutionary computing both models are
 +
possible and have their merits.
 +
Evolutionary robotics, where evolutionary algorithms are used to aid robot design, seems like a clear
 +
candidate for using learning and local search to improve efficiency. The analogy is very clear - each
 +
individual in the population represents a robot with a different control system ("brain") and perhaps
 +
also with a different body. A hybrid algorithm that applies some sort of local optimization to the robot
 +
during evaluation would be completely analogous to an animal improving by learning during its
 +
lifetime.
 +
There are several possible tasks availible related to lifetime learning, including
 +
* Comparing robot evolution with and without a lifetime learning process
 +
* Investigating how lifetime learning affects differences between simulation results and the real world
 +
* Implementing Lamarckian learning with an indirect encoding (i.e. performing phenotype to
 +
genotype translation)
 +
 +
In any case the project would include roughly the following components:
 +
* Investigation into provious research on evolutionary robotics and hybrid evolutionary algorithms,
 +
and then writing an essay on the topics.
 +
* Planning and implementation of relevant methods
 +
* Planning and implementation of suitable experiments to test these methods, in real life or in some
 +
kind of simulation.
 +
 +
== Supervisors ==
 +
 +
* [http://www.mn.uio.no/ifi/personer/vit/kyrrehg/ Kyrre Glette]
 +
* [http://www.mn.uio.no/ifi/personer/vit/eivinsam/ Eivind Samuelsen]
 +
 +
== Vil ha Table of contents ==

Current revision as of 11:03, 1 September 2014

Contents

Contact

Else-Line Ruud

  • Mail
  • Phone: 93851900

Description of master project

Lifetime Learning - Machine learning in an evolutionary robotics context In evolutionary computing, there are two main models for the application of local search or machine learning during the lifetime of an individual: Lamarckian learning, where the genes of an individual is updated so that lifetime experiences gets passed on to new generations, and Baldwinian learning, where only the fitness of an individual is changed. Baldwinian learning is closest to the the way the evolutionary process in nature is understood today, but in evolutionary computing both models are possible and have their merits. Evolutionary robotics, where evolutionary algorithms are used to aid robot design, seems like a clear candidate for using learning and local search to improve efficiency. The analogy is very clear - each individual in the population represents a robot with a different control system ("brain") and perhaps also with a different body. A hybrid algorithm that applies some sort of local optimization to the robot during evaluation would be completely analogous to an animal improving by learning during its lifetime. There are several possible tasks availible related to lifetime learning, including

  • Comparing robot evolution with and without a lifetime learning process
  • Investigating how lifetime learning affects differences between simulation results and the real world
  • Implementing Lamarckian learning with an indirect encoding (i.e. performing phenotype to

genotype translation)

In any case the project would include roughly the following components:

  • Investigation into provious research on evolutionary robotics and hybrid evolutionary algorithms,

and then writing an essay on the topics.

  • Planning and implementation of relevant methods
  • Planning and implementation of suitable experiments to test these methods, in real life or in some

kind of simulation.

Supervisors

Vil ha Table of contents

Personal tools
Front page