Coupling stochastic EM and approximate Bayesian computation for parameter inference in state-space models

Fast maximum likelihood estimation for general hierarchical models

Hierarchical statistical models are important in applied sciences because they capture complex relationships in data, especially when variables are related by space, time, sampling unit, or other shared features. Existing methods for maximum likelihood estimation that rely on Monte Carlo integration over latent variables, such as Monte Carlo Expectation Maximization (MCEM), suffer from drawbacks in efficiency and/or generality. We harness a connection between sampling-stepping iterations for such methods and stochastic gradient descent methods for non-hierarchical models: many noisier steps can do better than few cleaner steps. We call the resulting methods Hierarchical Model Stochastic Gradient Descent (HMSGD) and show that combining efficient, adaptive step-size algorithms with HMSGD yields efficiency gains. We introduce a one-dimensional sampling-based greedy line search for step-size determination. We implement these methods and conduct numerical experiments for a Gamma-Poisson mixture model, a generalized linear mixed models (GLMMs) with single and crossed random effects, and a multi-species ecological occupancy model with over 3000 latent variables. Our experiments show that the accelerated HMSGD methods provide faster convergence than commonly used methods and are robust to reasonable choices of MCMC sample size.

Two-Stage Bayesian Optimization for Scalable Inference in State-Space Models

State-space models (SSMs) are a rich class of dynamical models with a wide range of applications in economics, healthcare, computational biology, robotics, and more. Proper analysis, control, learning, and decision-making in dynamical systems modeled by SSMs depend on the accuracy of the inferred/learned model. Most of the existing inference techniques for SSMs are capable of dealing with very small systems, unable to be applied to most of the large-scale practical problems. Toward this, this article introduces a two-stage Bayesian optimization (BO) framework for scalable and efficient inference in SSMs. The proposed framework maps the original large parameter space to a reduced space, containing a small linear combination of the original space. This reduced space, which captures the most variability in the inference function (e.g., log likelihood or log a posteriori), is obtained by eigenvalue decomposition of the covariance of gradients of the inference function approximated by a particle filtering scheme. Then, an exponential reduction in the search space of parameters during the inference process is achieved through the proposed two-stage BO policy, where the solution of the first-stage BO policy in the reduced space specifies the search space of the second-stage BO in the original space. The proposed framework's accuracy and speed are demonstrated through several experiments, including real metagenomics data from a gut microbial community.

Spectral density-based and measure-preserving ABC for partially observed diffusion processes. An illustration on Hamiltonian SDEs

Approximate Bayesian computation (ABC) has become one of the major tools of likelihood-free statistical inference in complex mathematical models. Simultaneously, stochastic differential equations (SDEs) have developed to an established tool for modelling time-dependent, real-world phenomena with underlying random effects. When applying ABC to stochastic models, two major difficulties arise: First, the derivation of effective summary statistics and proper distances is particularly challenging, since simulations from the stochastic process under the same parameter configuration result in different trajectories. Second, exact simulation schemes to generate trajectories from the stochastic model are rarely available, requiring the derivation of suitable numerical methods for the synthetic data generation. To obtain summaries that are less sensitive to the intrinsic stochasticity of the model, we propose to build up the statistical method (e.g. the choice of the summary statistics) on the underlying structural properties of the model. Here, we focus on the existence of an invariant measure and we map the data to their estimated invariant density and invariant spectral density. Then, to ensure that these model properties are kept in the synthetic data generation, we adopt measure-preserving numerical splitting schemes. The derived property-based and measure-preserving ABC method is illustrated on the broad class of partially observed Hamiltonian type SDEs, both with simulated data and with real electroencephalography data. The derived summaries are particularly robust to the model simulation, and this fact, combined with the proposed reliable numerical scheme, yields accurate ABC inference. In contrast, the inference returned using standard numerical methods (Euler-Maruyama discretisation) fails. The proposed ingredients can be incorporated into any type of ABC algorithm and directly applied to all SDEs that are characterised by an invariant distribution and for which a measure-preserving numerical method can be derived.