Massive asynchronous master-worker EA for nuclear reactor optimization
Mathieu Muniglia, Sebástien Vérel, Jean-Charles Le Pallec, Jean-Michel Do · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2017
In the global goal to increase the part of the intermittent renewable energies in the French energy mix, the production of nuclear energy has to be adapted to face the power variations. We propose to optimize the main variables of the control rods of a nuclear power plant to improve its management, and increase the safety margins in case of a more heckled load-following schedule due to intermittent renewable energies. Using a multi-physics simulator, the criteria of interest can be computed in few minutes of computation, and we are thus facing a black-box combinatorial optimization problem with expensive evaluation. Hence, we propose a parallel asynchronous master-worker Evolutionary Algorithm scaling up to thousand computing units. From a practical optimization point of view, one main difficulty is the tuning of algorithm parameters such as mutation rates. In this work, we perform a fitness landscape analysis on this expensive real-world problem, and show that it is possible to tune the mutation parameters according to the low-cost estimation of the fitness landscape structure.