A generic distributed computing environment built for the internal
network of the European Space Agency has been developed and used
to distribute different global optimisation techniques (ACT-DC). In this project we use this environment to introduce a new distributed global multi-objective optimizer (DiGMO) that makes use of all the different available solvers to solve the same problem. First we try to assign to each client a randomly selected solver (here we deal with differential evolution, particle swarm, simulated annealing and genetic algorithm) and we already find a performance improvement with respect to each of the distributed stand-alone solver versions. We then try to apply some rules to learn the best combination of the available solving strategies. We find that a further performance improvment is possible if the probability to assign to each client a given solver depends on the performance history of the different solvers in the problem considered.
|
