Optimization based on dynamic and hybrid metaheuristics via DEVS simulation

International audienceIn this paper we present an efficient decision-making framework allowing optimization via simulation combining metaheuristics algorithms and DEVS formalism. A proposed object-oriented approach provides an universal interface between any existing DEVS models and some existing metaheuristics. To do this we explode the optimization algorithms into several actions involved in an optimization loop. Concerning the evaluation step which one can find in each metaheuristic we propose to externalize it into decision models using DEVS messages. This interconnection of components generates the following new concepts: (i) event-driven metaheuristic parametrization that allows an automatic execution of the associated algorithm; (ii) creation of dynamic and hydrid metaheuristics used for hard-optimization problems; (iii) decision support variables temporization that allows optimization on a specific time interval. We validated the effectiveness of the previous concepts through the implementation of the popular genetic algorithms and the more recent harmonic search algorithm. Furthermore the validation is completed on a real case example concerning medical treatment through a simple glycemic DEVS model. Simulation results show that the resulting treatement reduces the risk of over-tratement or under-treatement. 1. INTRODUCTION Modeling and simulation of complex systems are widely used in research and industry in order to study the behavior of systems. These studies include planning, prediction, prevision, computer-aided design, etc. tasks. To improve the performance of a studied system the modeling and simulation phases are integrated with an optimization technique. The simulation allows improving the purposes of a given model. The integration of optimization techniques into modeling and simulation relies on the evolution of the model using decisions based on previous simulation results. Usually a specialist of the domain performs the evolution manually. Therefore, it seems to be natural to try to find the set of parameters that optimizes the system performance