For the purpose of Market Consistent actuarial valuation, we need to calibrate our models to market data. Traditional optimization methods based only on gradient (Newton, quasi-Newton…) fail, or provide volatile results. Indeed, they search for a solution in a local neighborhood, while objective functions might exhibit multiple local minima. The R package mcGlobaloptim achieves global optimization, combining Monte Carlo and Quasi-Monte Carlo simulation with local searches. The local searches can be speeded-up by the use of your computer’s (multiple) cores.
Currently, the function multiStartoptim generates pseudo-random and quasi-random numbers within the search domain speciﬁed by its bounds. Local searches are then performed by a user-selected method, either on the whole set of numbers generated as starting points, or only on the points lying under the objective function’s median. When (and only when) a high number of local searches are to be performed, parallel computation can be used to speed-up the search