EvoNeuro and Sferes Documentation
Next

Documentation 1.0

Copyright © 2011

CeCiLL

Sferes 2

Introduction
Main features
Authors
Where to go next?
Specific sections for ISIR members

Sferes2 is a framework for evolutionary computation (EC) experiments and especially for evolutionary robotics (ER). Its main goal is to help researchers in EC and ER to efficiently try new ideas. Sferes2 has been inspired by sferes http://sferes.lip6.fr, another (older and no more maintained) framework for ER.

Sferes2 allows to design efficient programs on modern (multicore) computers: our experiments typically require more than 24 hours of computation; in this situation it can be profitable to trade some complexity of implementation for some hours of computation. Nevertheless, we try to keep the framework as simple as possible. More specifically, the following choices were made in the initial design:

  • use of modern c++ techniques (template-based programming) to employ object-oriented programming without the cost of virtual functions;
  • use of Intel TBB to take full advantages of multicore and SMP systems;
  • use of boost libraries when it's useful (shared_ptr, serialization, filesystem, test,...);
  • use of MPI to distribute the computational cost on clusters;
  • a full set of unit tests;
  • no configuration file: a fully optimized executable is built for each particular experiment.

Appart from this documentation, information about Sferes 2 can be found in Sferes_v2: Evolvin' in the Multi-Core World. If you use Sferes 2 in your work, please cite the previous publication.

The source code of many experiments using sferes is available on the evorob_db website

This documentation assumes that the reader is reasonably familiar with evolutionary algorithms. For references you can check the following books about evolutionary computation:

  • Deb, K. (2001). Multi-Objective Optimization Using Evolutionary Algorithms. Wiley.
  • De Jong, K. A. (2002). Evolutionary Computation. The MIT Press.

Sferes2 currently includes the following EA and toolboxes (this is a non-exhaustive list which contains only the most stable modules):

  • EA

    • a rank-based single-objective EA
    • CMAES (Covariance Matrix Adaptation Evolution Strategie)
    • NSGA-II (multiobjective EA)
    • Epsilon-MOEA (multiobjective EA)

  • Genotypes / Phenotypes:

    • Evolution of bit strings
    • Evolution of real parameters using different operators (SBX, gaussian, etc)

  • Simulators

    • (optional) mobile robot (Khepera-like)
    • (optional) cartpole (single and double)

  • Other

    • (optional) evolution of neural-networks

Download Downloading sferes2

Compilation How to compile sferes2 on your system

Tutorial/Example A tutorial/example with explanation about the general structure of sferes2 experiment files, the example file is generated automatically in the next tutorials.

Neurobiology Tutorial I A tutorial to create an experiment using the neural network module of sferes2, this enables you to create a neural network and evolve its weights and parameters.

Neurobiology Tutorial II A tutorial to evolve the topology of a network in sferes2

Neurobiology Tutorial III A tutorial to evolve a network of maps using the EvoNeuro module from sferes II.

Reference A reference manual / guide to the source code

Torque Introduction A reference to launching experiments on the ISIR cluster

Next