Simulation model calibration with correlated knowledge-gradients

We address the problem of calibrating an approximate dynamic programming model, where we need to find a vector of parameters to produce the best fit of the model against historical data. The problem requires adaptively choosing the sequence of parameter settings on which to run the model, where each run of the model requires approximately twelve hours of CPU time and produces noisy non-stationary output. We describe an application of the knowledge-gradient algorithm with correlated beliefs to this problem and show that this algorithm finds a good parameter vector out of a population of one thousand with only three runs of the model.