What is SFERES ?

SFERES is a framework that gathers an evolution framework with a simulation framework. It is dedicated to experiments involving the design of artificial agents through artificial evolutionary algorithms, like Genetic Algorithms or Evolution Strategies, for instance. Each experiment is described by a simple XML file that defines each part of the experiment: which selection algorithm to use, which genotype (bitstring, vector of real values or more sophisticated structures), which phenotype to control an agent (adhoc controller, neural network...), what kind of agent (wheeled or legged robot for instance), what fitness ... SFERES is intended to provide a development environment that facilitates the implementation of new algorithms by focusing only on the new parts while reusing already developed components. Furthermore, as other components are available, it becomes easier to compare the performance of algorithms in exactly the same conditions.

What can SFERES do ?

It is clearly dedicated to the evolution of control architectures for simulated agents. We have used it to evolve neural networks for the low level control of blimps, helicopters, flapping wing aircraft or other kind of robots. It can also be used to evolve a robot morphology.

We have actually used the simulation part alone in some of our experiments, as it allowed us to keep a same development environment and has facilitated the integration of different part of a whole controller that has not been entirely evolved.

SFERES can also be used for experiments involving evolution but without a simulation, but in this case, it will loose one of its main interesting features. Anyway, some algorithms developed in the context of artificial life can be directly tested on a simple function optimization, for instance, without the need to reimplement it.

Summary of SFERES features:

• evolution and visualization tool
• XML configuration files
• Compressed context file
• Statistics of an experiment: best individual, average performance, worst individual and users can add new statistics to measure datas specific to a particular component
• Advanced debugging library included (libcwd)
• Scheduler of discrete evenements (unique or repetitive) describing an evaluation
• Simple or multi-agent experiments