Abstract
Our paper deals with inferring simulator-based statistical models given some observed data. A simulator-based model is a parametrized mechanism which specifies how data are generated. It is thus also referred to as generative model. We assume that only a finite number of parameters are of interest and allow the generative process to be very general; it may be a noisy nonlinear dynamical system with an unrestricted number of hidden variables. This weak assumption is useful for devising realistic models but it renders statistical inference very difficult. The main challenge is the intractability of the likelihood function. Several likelihood-free inference methods have been proposed which share the basic idea of identifying the parameters by finding values for which the discrepancy between simulated and observed data is small. A major obstacle to using these methods is their computational cost. The cost is largely due to the need to repeatedly simulate data sets and the lack of knowledge about how the parameters affect the discrepancy. We propose a strategy which combines probabilistic modeling of the discrepancy with optimization to facilitate likelihood-free inference. The strategy is implemented using Bayesian optimization and is shown to accelerate the inference through a reduction in the number of required simulations by several orders of magnitude.
Introduction
In scientific disciplines ranging from ecology to epidemiology, researchers frequently employ "simulator-based models." These generative models are computer programs that produce synthetic data based on a set of input parameters ($\theta$). The primary objective is to identify the parameters that yield outputs most closely resembling real-world observations.
However, these models often possess an intractable likelihood. There is no closed-form mathematical expression for the probability of the data given the parameters. Consequently, one can only execute the simulator and observe the outcomes. Traditional approaches, such as Approximate Bayesian Computation (ABC) or "Synthetic Likelihood," rely on computationally intensive brute-force methods, simulating millions of datasets to find matches. This becomes impractical when individual simulations are time-consuming.
Methodology: The BOLFI Framework
The proposed BOLFI (Bayesian Optimization for Likelihood-Free Inference) framework shifts from sample-average approximations to regression-based approximations. Instead of treating simulations as independent events, BOLFI models the relationship between parameters and data discrepancy as a smooth function.
- Gaussian Process (GP) Surrogate: A GP is used to model the discrepancy—the "distance" between simulated and observed data. This serves as a probabilistic emulator of the simulator.
- Bayesian Optimization (BO): Rather than random sampling, BOLFI uses an acquisition function (e.g., Lower Confidence Bound) to determine the next parameter set to simulate. This balances exploration of unknown regions with exploitation of areas where discrepancy is low.
- Likelihood Derivation: Once the GP learns the discrepancy function's shape, an approximate likelihood and posterior distribution can be derived analytically without further simulations.
A key insight is that the discrepancy function's smoothness allows BOLFI to "share" information across the parameter space. By focusing on promising regions, BOLFI avoids wasting resources on "dead zones" that clearly do not match the data.
Experimental Evaluation
The effectiveness of BOLFI was evaluated on three models, including the chaotic Ricker model and a complex model of bacterial infections in day care centers.
Key Results
The efficiency gains provided by BOLFI are substantial:
- Ricker Model: BOLFI achieved results comparable to Markov Chain Monte Carlo (MCMC) while requiring 600 times fewer simulator calls.
- Bacterial Infection Model: BOLFI produced an informative posterior with only 50 simulations per parameter dimension. Compared to standard methods requiring 15,000 simulations, this represents a 300x speedup.
- Computational Time: For the bacterial model, traditional PMC-ABC took 4.5 days on a 200-core cluster, whereas BOLFI completed in approximately 1.5 hours on a single desktop.
Discussion and Future Work
While BOLFI offers significant acceleration, challenges remain. Gaussian Processes can struggle with high-dimensional parameter spaces (typically $d > 20$). Future work may involve integrating high-dimensional BO techniques like random embeddings. Additionally, because BOLFI focuses on the "mode" of the distribution for efficiency, it may be less precise in estimating the "tails" (extreme parameter values). Refining acquisition rules for the noisy discrepancies inherent in likelihood-free inference is another promising avenue for research.
Summary updated