Chris Jewell, Lancaster University
Speaker: Chris Jewell, Lancaster University
Title: gemlib: probabilistic programming for state transition models
Abstract:
State-transition models are central to applications in epidemiology and ecology, yet statistical inference remains challenging due to high-
dimensional latent state spaces, strong temporal dependence, and
intractable likelihoods. Existing probabilistic programming frameworks
are limited in capturing the semantics of such models, with the result
that the software ecosystem is dominated by tightly coupled, model-
specific packages that hinder reuse and extension.
Here, we present "gemlib", an attempt at studying how state-transition
models decompose into units that mimic the epidemiologists thought
process when defining a model. We take a mathematical approach to
separating the concerns of state-transition specification from the
integration routine (ODE or discrete- or continuous-time stochastic),
such that models may be reasoned about, and modified quickly without
introducing bugs. Using a similar decompositional approach, we show
how a MCMC framework may be constructed using a functorial design,
implying an obvious software architecture and MCMC algorithm grammar specification.
Though "gemlib" is implemented in Python, our approach uses
denotational semantics allowing straightforward re-implementation in
any Turing-complete language. Our hope is to inspire a conversation
not about which Turing-complete language is best for modelling, but
rather how models and their associated algorithms may be formalised
into a well-understood calculus that increases the transparency,
reliability, and reproducibility of our field.