kayros 0.1.0

KAYROS — exact & anytime solver for duration-minimization time-dependent vehicle routing (TDVRPTW / TDVRP)

KAYROS

KAYROS is an exact & anytime solver for duration-minimization time-dependent vehicle routing problems — TDVRPTW (with time windows) and TDVRP — benchmarked on the canonical MAMUT-routing TD instance families.

Status: alpha. v0.1.0 ships an anytime time-dependent Ant Colony Optimization heuristic on an exact non-decreasing continuous piecewise-linear (NDCPWLF) arrival-time engine, benchmarked on all four MAMUT TD families (it seeded the store's initial best-known solutions at scale). The time-dependent local-search layer is next. See the roadmap below.

Design principles

• One command install, no proprietary dependency. The default build is pure open source (C++23 + pybind11). The optional exact branch-price-and-cut component of Lera-Romero et al. (requires CPLEX) is strictly opt-in (-DKAYROS_WITH_LERA=ON, source build only) and never ships in wheels.
• Exact arithmetic, checker-refereed. Route durations are computed on NDCPWLF arrival-time functions with exact doubles — no epsilon comparisons. Every solution kayros reports is priced by the reference checker of mamut-routing-lib; the checker's value is the value.
• POD core. The C++ core is plain structs, flat arrays and free functions — optimization-kernel style, no framework.

Install

pip install kayros

This pulls everything, including the benchmark loaders and the reference checker (mamut-routing-lib). For development:

git clone https://github.com/0nyr/kayros && cd kayros
pip install -e . --group dev    # pip >= 25.1 (or: uv pip install -e . --group dev)

Requirements: Python ≥ 3.11; building from source (sdist or checkout) additionally needs a C++23 compiler and CMake ≥ 3.26 (fetched automatically by the build backend when missing).

Usage

import kayros

# Any MAMUT-routing TD instance (.vrp.json with its .atf.json sidecar next to it)
instance_path = "benchmarks/TDVRPTW/Dabia2013/n=25/C101.vrp.json"
solution = kayros.solve(instance_path, time_limit=10.0, seed=42)
print(solution.duration, solution.num_routes, solution.status)

# Anytime: react to every new incumbent while the solve keeps running
def on_incumbent(incumbent, routes):
    print(f"[{incumbent.seconds:7.2f}s] {incumbent.value:.6f} ({incumbent.origin})")

solution = kayros.solve(instance_path, time_limit=60.0, on_incumbent=on_incumbent)

# Feed the MAMUT BKS pipeline: solution.to_benchmark_solution() returns the
# artifact accepted by mamut_routing_lib.td.bks.save_td_solution_as_bks_if_improved.

solution.duration is always the value computed by the reference checker (mamut_routing_lib.td.check_td_solution) — never an internal approximation.

Roadmap (stage 1)

• M3.0 — package scaffold, CI, PyPI wiring
• M3.1 — NDCPWLF composition engine + POD instance/route core
• M3.2 — exact equivalence gate against the reference checker (Dabia2013): 513 tests, 336 instances, zero divergences
• M3.3 — kayros.solve(): greedy construction + TD-ACO
• M3.4 — all four MAMUT TD families × {TDVRPTW, TDVRP}
• M3.5 — large-scale runs on Grid'5000: seeded the initial best-known solutions for all 1352 MAMUT TD instances
• M3.6 — anytime API (on_incumbent, time budgets) + v0.1.0 on PyPI
• M3.7 — time-dependent local search layer (LCA-BST move evaluation, Blauth et al. 2024)
• Later: optional exact BPC (kayros[lera])

Provenance

KAYROS is developed by Florian Rascoussier (Onyr) as part of a PhD in operations research (IMT Atlantique / INSA Lyon), under the supervision of Romain Billot, Christine Solnon and Lina Fahed. The NDCPWLF composition engine follows Visser & Spliet (2020)'s move-evaluation theorems; the TD-ACO is a rewrite of the author's heuristic layer originally built on the TDVRPTW solver of Lera-Romero, Rönnqvist & Ljungqvist (2020, MIT-licensed). MIT license.