Tools: Neural Network Toolbox (ISEE)*
The ISEE (Integrated Structure Evolution Environment) is a powerful software
platform for the evolution of recurrent neural networks (RNN) for behaviour
control of autonomous agents. It was originally developed to connect the
ENS⊃3
algorithm with various robot simulators as fast and easy as possible. The
connection to other evolution software is also possible. This ISEE platform
combines three different components:
- the evolutionary core program (EvoSun),
- the simulator and
- the general interface (HINTON).
The EvoSun program implements variation, evaluation and selection operators
acting on populations of (recurrent) neural networks. For application in the
field of evolutionary robotics the evaluation of a controlling neural network is
done during the simulation of the robot in a defined environment.
To keep the ISEE platform open to different robot platforms the HINTON-component
is developed as a general interface. It realizes the interchange of evolved
networks between the evolution core (EvoSun) and the robot simulator.
In addition, the HINTON-interface allows evolving several populations in
parallel, i.e. the ISEE platform is easy to use for co-evolution.
Furthermore, parallel evolution of one population can be easily done. This type
of distributed evolution can bring an important speed up of the evolutionary
process.
Last but not least, the interface HINTON is also ready to be used as stand
alone. In this mode different control architectures can be directly tested
without a difficult and time consuming implementation in the simulator.
Contact:
Frank Pasemann
and
Keyan Ghazi-Zahedi
* This software was developed at the former Intelligent Dynamics (INDY) group at
the Fraunhofer Institute AIS (now FhI-IAIS).
|