Genetic algorithms imitate the collective learning paradigm found in living nature. They derive their power largely from their implicit parallelism gained by processing a population of points in the search space simultaneously. In this paper, we describe an extension of genetic algorithms making them also explicitly parallel. The advantages of the introduction of a population structure are twofold: firstly, we specify an algorithm which uses only local rules and local data making it massively parallel with an observed linear speedup on a transputer-based parallel system, and secondly, our simulations show that both convergence speed and final quality are improved in comparison to a genetic algorithm without population structure.