Parameter estimation/calibration is one of the key challenges when working with complex simulators. Approximate Bayesian computation (ABC) algorithms (and the related class of methods known as history matching) are Monte Carlo algorithms that seek to find regions of parameter space that produce reasonable agreement between the simulator prediction and the data. Naive implementations of ABC require more simulation that is typically possible and so to overcome this cost, it is common to make use of a surrogate model, or emulator, which approximates the simulator output. Design for building emulators of computer experiments usually focusses on how to build accurate global approximations. However, if our aim is calibration, this can be wasteful.
In this talk I will discuss the use of emulators within ABC/history-matching, and describe design strategies that can be used, particularly entropy based designs that minimize the uncertainty in the classification surface.