Stochastic simulations are widely used in application domains where no deterministic laws can be used to predict the futur state of the system.
The results of such simulations being variable by nature, it is then mandatory to replicate the simulation execution in order to retreive exeprimentally the distribution of possible eventes and then obtain their probabilities. In the general case, the Monte-Carlo method is classically used. Nevertheless, if computing resources are limited or targeted events are rare, MC can be inefficient.
In this thesis, we introduce a new simulation execution policy based on the clustering, selection and cloning of replications states, in order to improve the result quality for a fixed computing cost while being as generic as possible. The general idea of this policy is to constraint the evolution of the replications by periodically selecting and cloning those potentially leading to interesting results for the decision maker.
The policy is validated on simple academic agent-based models (prey predator, and virus transmission), then experimented on a more complex traffic simulation. A probabilistic variant of the policy is finally presented.
(*) The ELSAT2020 project is co-financed by the European Union with the European Regional Development Fund, the French state and the Hauts de France Region Council.