HYPERPARAMETER OPTIMIZATION FOR APPROXIMATE BAYESIAN COMPUTATION

Approximate Bayesian computation is a popular methodology for simulation-based parameter inference in scenarios where the likelihood function is either analytically intractable or computationally infeasible. The likelihood of simulator parameters fitting given data is approximated by iteratively simulating samples that are generated according to a specified prior distribution. The convergence speed and the quality of inference are highly sensitive to the choice of hyperparameters such as the chosen summary statistic, the value of the acceptance threshold and the distance function. The choice is typically left to the domain expert as summary statistics vary across disciplines and threshold values are problem-specific. This work explores automated hyperparameter optimization for approximate Bayesian computation using Bayesian optimization as an alternative to time consuming manual selection. The problem setting assumes availability of a fast low-fidelity simulator, which is used during the optimization process. The optimized hyperparameters are then used to perform inference using the high-fidelity simulator.