# Fitness functions for low-dimensional parts of cooperative coevolution algorithms

In cooperative coevolution algorithms, a high dimensional vector is broken into smaller vectors, each of which is optimized separately using EAs for fewer dimensions and then recombined.

What is the fitness function used to optimize those? It can't be the same benchmark function, as those are for higher dimensions.

I am specifically talking with respect to the algorithm DECC-G ("Large Scale Evolutionary Optimization Using Cooperative Coevolution" by Z. Yang, K. Tang, X. Yao). What is the fitness function used to optimize subpop in SaNSDE and wpop in DE?

## 1 Answer

Every sub-vector is evaluated by concatenating it with the best-fit individuals (sub-vectors) from the rest of the sub-populations to form a so called context vector.

The context vector contains all the parameters required by the benchmark function and is fed into it for fitness evaluation.

This is actually where the cooperation happens.

DECC-G is "just" a cooperative co-evolutionary (CC) scheme based on Differential Evolution (DE) and a random split/grouping of the vectors/individuals (-G).

It shares with the other co-evolutionary algorithms the general architecture:

(Coevolutionary model of three species shown from the perspective of each in turn. Image from Cooperative Coevolution: An Architecture for Evolving Coadapted Subcomponents by Mitchell A. Potter, Kenneth A. De Jong)

As a side note the reference implementation of DECC-G can be found here.