Coevolution

Coevolution in Evolutionary Optimization

How does coevolution affect evolutionary optimization?

The coevolution project is a rather recent addition to the overall EvCA research program.
Below is a short overview of this project.
Please see the EvCA papers on coevolution for more details about this project.

Coevolutionary Models

Evolutionary optimization models generally utilize a static selection function; individuals are selected on the basis of how well they comply to the objective function. In a coevolutionary model the success of individuals (i.e. their number of offspring) depends on circumstances that represent only a subset of the objective function. The resulting evolutionary dynamics may lead to continuous evolutionary change, i.e. red queen dynamics, or speciation of the populations, or to the evolution of general behavior, i.e. optimization of individual behavior. The sparseness of the fitness evaluation, i.e. individuals are only evaluated on a subset of all fitness cases, may have as side-effects an increased efficiency in the optimization process, it gives more freedom to evolve with respect to the complete objective funtion, and it may lead to a continuous effective selection gradient.

Coevolving Cellular Automata

In the EvCA group we study coevolutionary dynamics in evolutionary optimization models in order to study some of the mentioned ideas. The objective function is the density classification function for cellular automata. The coevolution is defined between the cellular automata (CAs) and initial conditions (ICs). Whereas in the standard evolutionary optimization process CAs are evaluated on the basis of a large set of ICs which are randomly created every time step, in the coevolutionary model the ICs coevolve with the CAs. CAs are selected on the basis of how well they classify ICs, whereas ICs are selected on the basis of how difficult they are to be classified correctly by the CAs; CAs and ICs have antagonistic selection pressures.
We want to study if, and how, coevolution can be beneficial for evolutionary optimization processes. In this context the (maintenance of) diversity in the populations seem of crucial importance. We study different mechanisms that promote diversity in evolving populations, e.g. sharing methods and spatial embedding, i.e. parallel coevolution, of the evolutionary process. We also study if eco-evolutionary dynamics, i.e. combined ecological and evolutionary dynamics, can lead to evolutionary processes that have self-organizing characteristics that lead to problem independent optimization.