Last friday we were discussing the issues of decentralization and multiobjective optimization in Genetic Algorithms, in more detail:
- Decentralizing GAs is a radical way of parallelizing GAs in which the parallel infraestructure is devoid of any central server. Hence, individuals have no global but local information provided by the neighbourhood.
- Multiobjective GAs are GAs optimizing several conflicting objectives. In such a way the output of the GA is not a single but a set of solutions (usually named pareto-front) and it is based on some criterion of dominance.
Matching both issues is not straightforward since inserting a new solution within the set of solutions requires of a global comparison through the pareto-front and, talking about a decentralized GA, the pareto-front should be decentralized in advance. The consequence is that there are few works tackling such challenge. One of them is the paper A Cellular Genetic Algorithm for Multiobjective Optimization by Nebro et al.which can be considered as an step towards, although not a definitive solution to decentralized multiobjective optimization. The authors propose:
- A decentralized GA: the Cellular Genetic Algorithm in which every individual is a single process.
- An external archive (and centralized) to store the non-dominated solutions.
- A simulation of the parallel environment.
In spite of the drawback of having a centralized archive which does not allow a real parallel/decentralized execution, results show to be algorithmically competitive against the state of the art algorithms NSGA-II and SPEA2. Therefore, one of the possible reading from the paper is that decentralized GAs are promising, although there is no a pure decentralized GA for Multiobjective optimization yet.