Browsing by Author "Kokko, Jan"

In this thesis we present a new likelihood-free inference method for simulator-based models. A simulator-based model is a stochastic mechanism that specifies how data are generated. Simulator-based models can be as complex as needed, but they must allow exact sampling. One common difficulty with simulator-based models is that learning model parameters from observed data is generally challenging, because the likelihood function is typically intractable. Thus, traditional likelihood-based Bayesian inference is not applicable. Several likelihood-free inference methods have been developed to perform inference when a likelihood function is not available. One popular approach is approximate Bayesian computation (ABC), which relies on the fundamental principle of identifying parameter values for which summary statistics of simulated data are close to those of observed data. However, traditional ABC methods tend have high computational cost. The cost is largely due to the need to repeatedly simulate data sets, and the absence of knowledge of how to specify the discrepancy between the simulated and observed data. We consider speeding up the earlier method likelihood-free inference by ratio estimation (LFIRE) by replacing the computationally intensive grid evaluation with Bayesian optimization. The earlier method is an alternative to ABC that relies on transforming the original likelihood-free inference problem into a classification problem that can be solved using machine learning. This method is able to overcome two traditional difficulties with ABC: it avoids using a threshold value that controls the trade-off between computational and statistical efficiency, and combats the curse of dimensionality by offering an automatic selection of relevant summary statistics when using a large number of candidates. Finally, we measure the computational and statistical efficiency of the new method by applying it to three different real-world time series models with intractable likelihood functions. We demonstrate that the proposed method can reduce the computational cost by some orders of magnitude while the statistical efficiency remains comparable to the earlier method.