10.2.2.2. TEgoModeler
There is only one modeler currently available TEgoKBModeler
To deal with solutions that are currently under evaluation by the cpu resources (asynchronous parallelism), this modeler uses the kriging believe principle (it trusts in the model prediction). Two models are created: a first one, built with all the evaluated solutions, is used to estimate solutions under evaluation; a second model, built with both evaluated solutions and estimated ongoing solutions, is used by the solver to find the next solution to evaluate. Predictions is not significantly affected in the second model but the variances are, especially around ongoing solutions. The EI objective takes it in account, naturally driving next solutions away from them.
As it is an optimization, advancing in its search, EGO will generate solution near existing ones. It is a difficulty for model construction that can leads to Cholesky decomposition errors. To get around this problem, users can use the kriging regulation.
it uses