MULTILEVEL MONTE CARLO METAMODELING

Multilevel Monte Carlo (MLMC) methods have been used by the information-based complexity community in order to improve the computational efficiency of parametric integration. We extend this approach by relaxing the assumptions on differentiability of the simulation output. Relaxing the assumption on the differentiability of the simulation output makes the MLMC method more widely applicable to stochastic simulation metamodeling problems in industrial engineering. The proposed scheme uses a sequential experiment design which allocates effort unevenly among design points in order to increase its efficiency. The procedure's efficiency is tested on an example of option pricing in the Black-Scholes model.