Roadmap

This document describes the planned direction for givp over the next twelve months. Items are subject to change based on user feedback and available contributor time.

Current version

v0.3.x — stable, feature-complete implementation of the GRASP + ILS + VND + Path Relinking metaheuristic for continuous black-box optimization.

Short-term (next 3 months)

• Parallel neighbourhood evaluation: allow users to provide a parallelism hint so VND neighbourhoods can be evaluated concurrently on multi-core machines.
• Progress callback: expose an optional callback parameter that receives the best solution found after each iteration, enabling custom early-stopping and progress bars.
• Expanded examples: add worked examples for combinatorial problems (e.g., TSP-style discretised objective) and multi-objective scalarisation.

Medium-term (3–6 months)

• Warm-start initialisation: allow callers to seed the elite pool with known-good solutions to accelerate convergence.
• Configurable path-relinking strategies: expose forward, backward, and random PR direction as an explicit option.
• Documentation improvements: add a dedicated Architecture page and benchmark comparison charts.

Long-term (6–12 months)

• Optional scikit-learn integration: expose givp as a scikit-learn-compatible BaseEstimator for hyper-parameter tuning workflows.
• Type-safe bounds specification: accept named-parameter bounds via a mapping in addition to the current sequence-of-pairs format.
• Async support: explore asyncio-compatible runner for use in Jupyter and async frameworks.

Out of scope

The following are explicitly out of scope for this project:

• Gradient-based optimisation — use SciPy or PyTorch for that.
• Integer / mixed-integer programming — dedicated MIP solvers (e.g., PuLP, OR-Tools) are better suited.
• GPU acceleration — not currently planned.

Feedback

If a feature you need is missing, please open a GitHub Issue with the label enhancement.