Rapid Bayesian Inference for Expensive Stochastic Models

Almost all fields of science rely upon statistical inference to estimate unknown parameters in theoretical and computational models. While the performance modern computer hardware continues grow, requirements for simulation models are growing even faster. This is largely due increase model complexity, often including stochastic dynamics, that necessary describe characterize phenomena observed using modern, high resolution, experimental techniques. Such rarely analytically tractable, meaning extremely large numbers simulations required parameter inference. In such cases, can be practically impossible. this work, we present new Bayesian techniques accelerate expensive by computationally inexpensive approximations inform feasible regions space, through learning transforms adjust biased approximate inferences closer represent correct under model. Using topical examples from ecology cell biology, demonstrate a speed improvement an order magnitude without any loss accuracy. represents substantial over current state-of-the-art methods computations when appropriate available. Supplementary files article available online.