Changelog¶

All notable changes to this project will be documented in this file.

[0.3.0] - 2026-01-25¶

Python package renamed from chronopt to diffid
Ask-Tell interface for optimisers and samplers, enabling stateful step-by-step execution
Optimisers unified into an Optimiser enum with shared run() API
ParameterRange type with improved Bounds API supporting unbounded parameters
Python bindings enhanced with magic methods (__repr__, __str__, etc.) for better REPL experience
Cached Diffsol parallelisation with single solver build per thread and improved multithreaded error management
Improved error hierarchy with EvaluationError, TellError, and ProblemBuilderError distinctions
Added MCMC proposals for Dynamic Nested Sampling with corresponding benchmarks

Result attributes renamed: sigma0 → step_size, standardised to evaluations and iterations
Codebase aligned to British English spelling conventions
Samplers refactored with Ask-Tell interface, matching optimiser patterns
Module restructure: types.rs moved to library level, error types consolidated
ScalarEvaluation and GradientEvaluation now use TryFrom trait implementations
Objective trait extended with has_gradient() method
Optimiser .tell() and .run() methods use relaxed error management

Dynamic Nested Sampling (DNS) sampler with Rust and Python bindings, including a dedicated scheduler, proposal generator, result struct, and corresponding tests and stubs.
Support for sensitivities in DiffsolProblem across dense and sparse backends, exposing gradient support through CostMetric and problem.evaluate_with_grad(...).
Problem-level parallel evaluation control, parallel execution support for DNS, and a time attribute on sampler results, with parallel sampler tests.
Adds ADAM optimisation algorithm with Python bindings and stub support.
Multi-cost / vectored cost support, including cost weighting utilities for composing and weighting multiple metrics.
Diffsol backend configuration updated to use LLVM-18 for sensitivity calculations and improved performance.
Adds benchmarking suite for Diffsol, Optimiser, and Samplers via cargo bench

Refactor DiffsolProblem's evaluate function for reduced clone's and aligned error management
Several robustness fixes to Diffsol integration, including catching panics during failure cases and rebuilding with penalties.
CMA-ES bug fixes and bound handling improvements, with associated tests.
Corrected PyProblem.evaluate gradient argument handling on the Python side.
Miscellaneous clippy warnings and CI-related issues.

Vector-valued optimisation via VectorProblemBuilder and associated Problem type, alongside renamed ScalarProblemBuilder for scalar objectives.
build methods on problem builders no longer consume the builder, enabling repeated build calls and multi-build workflows.

New dynamic nested sampling and model evidence examples with updated evidence values using more reasonable live point counts.
New bicycle model identification example with model selection via evidence.
Expanded predator–prey example suite, including a dedicated subdirectory comparing solver backends and a Diffeqpy-based identification workflow.
Updated JAX predator–prey example to use Diffrax.

Initial Chronopt release combining the Rust optimisation core with Python bindings via PyO3 and maturin packaging.
Diffsol integration with configurable dense and sparse backends, gradient support, and builder APIs for solver configuration.
Suite of optimisation algorithms including Nelder–Mead, CMA-ES, patience-based halting, and parallel evaluation helpers.
Cost metric implementations (SSE, RMSE, NLL) with Python-facing builder helpers and validation.
Sampling module with the initial Metropolis–Hastings implementation and worked Python examples (e.g., bouncy ball, predator–prey, Lotka–Volterra).
Extensive unit and integration tests covering builders, diffsol backends, sampling, and mathematical benchmarking suites.
Automated stub generation helpers, documentation updates, and GitHub Actions workflows for CI and trusted-publisher PyPI releases.