Abstract
Modular robots oer an important benet in evolutionary
robotics, which is to quickly evaluate evolved morphologies and control
systems in reality. However, articial evolution of simulated modular
robotics is a dicult and time consuming task requiring signicant computational
power. While articial evolution in virtual creatures has made
use of powerful generative encodings, here we investigate how a generative
encoding and direct encoding compare for the evolution of locomotion
in modular robots when the number of robotic modules changes.
Simulating less modules would decrease the size of the genome of a direct
encoding while the size of the genome of the implemented generative
encoding stays the same. We found that the generative encoding is signi
cantly more ecient in creating robot phenotypes in the initial stages
of evolution when simulating a maximum of 5, 10, and 20 modules. This
not only conrms that generative encodings lead to decent performance
more quickly, but also that when simulating just a few modules a generative
encoding is more powerful than a direct encoding for creating
robotic structures. Over longer evolutionary time, the dierence between
the encodings no longer becomes statistically signicant. This leads us to
speculate that a combined approach { starting with a generative encoding
and later implementing a direct encoding { can lead to more ecient
evolved designs.
robotics, which is to quickly evaluate evolved morphologies and control
systems in reality. However, articial evolution of simulated modular
robotics is a dicult and time consuming task requiring signicant computational
power. While articial evolution in virtual creatures has made
use of powerful generative encodings, here we investigate how a generative
encoding and direct encoding compare for the evolution of locomotion
in modular robots when the number of robotic modules changes.
Simulating less modules would decrease the size of the genome of a direct
encoding while the size of the genome of the implemented generative
encoding stays the same. We found that the generative encoding is signi
cantly more ecient in creating robot phenotypes in the initial stages
of evolution when simulating a maximum of 5, 10, and 20 modules. This
not only conrms that generative encodings lead to decent performance
more quickly, but also that when simulating just a few modules a generative
encoding is more powerful than a direct encoding for creating
robotic structures. Over longer evolutionary time, the dierence between
the encodings no longer becomes statistically signicant. This leads us to
speculate that a combined approach { starting with a generative encoding
and later implementing a direct encoding { can lead to more ecient
evolved designs.
| Original language | English |
|---|---|
| Title of host publication | Applications of Evolutionary Computation : 20th European Conference, EvoApplications 2017, Amsterdam, The Netherlands, April 19-21, 2017, Proceedings, Part I |
| Number of pages | 16 |
| Publisher | Springer |
| Publication date | 25 Jan 2017 |
| ISBN (Print) | 978-3-319-55848-6 |
| ISBN (Electronic) | 978-3-319-55849-3 |
| DOIs | |
| Publication status | Published - 25 Jan 2017 |
| Event | Evostar 2017: The Leading European Event on Bio-Inspired Computation - The Bazel, Vijzelstraat 32, Amsterdam, Netherlands Duration: 19 Apr 2017 → 21 Apr 2017 http://www.evostar.org/2017/ |
Conference
| Conference | Evostar 2017 |
|---|---|
| Location | The Bazel, Vijzelstraat 32 |
| Country/Territory | Netherlands |
| City | Amsterdam |
| Period | 19/04/2017 → 21/04/2017 |
| Internet address |
| Series | Lecture Notes in Computer Science |
|---|---|
| Volume | 10199 |
| ISSN | 0302-9743 |
Keywords
- Modular Robots
- Evolutionary Algorithms
- Direct & Generative Encodings