v5.0.1

What's Changed

• Add pre-commit + isort by @leonlan in #112
• Example notebook PFSP by @leonlan in #110
• Github templates, contributing text by @N-Wouda in #115

Full Changelog: v5.0.0...v5.0.1

v5.0.0

What's Changed

• Textual improvements by @leonlan in #89
• Rename weights to select by @leonlan in #96
• Add standalone documentation by @N-Wouda in #95
• Improve TSP example notebook by @N-Wouda in #101
• New selection scheme interface by @leonlan in #98
• New CVRP example notebook by @leonlan in #100
• alpha-UCB bandit operator selection scheme by @N-Wouda in #102
• Random operator selection scheme by @N-Wouda in #103
• Implement more callbacks by @N-Wouda in #104

Full Changelog: v4.1.0...v5.0.0

Version 4.1.0

What's Changed

• NoImprovement stopping criterion by @leonlan in #77
• RandomWalk acceptance criterion by @leonlan in #79
• WorseAccept acceptance criterion by @leonlan in #78
• ThresholdAcceptance acceptance criterion by @leonlan in #80
• GreatDeluge acceptance criterion by @leonlan in #81
• HillClimbing acceptance criterion by @leonlan in #82

Full Changelog: v4.0.0...v4.1.0

Version 4.0.0

Major version bump because of breaking changes introduced in the WeightScheme's interface, and its dependents.

Changelog

Issue #57 (PR #70, thanks @leonlan) introduces the optional only_after keyword argument to ALNS.add_repair_operator. This keyword argument can be used to indicate that a repair operator only works with a specific set of destroy operators. One can use this argument as follows:

from alns import ALNS

def destroy_op1(current, rnd_state):
  pass

def destroy_op2(current, rnd_state):
  pass

def repair_op1(destroyed, rnd_state):
  pass

def repair_op2(destroyed, rnd_state):
  pass

alns = ALNS()
alns.add_destroy_operator(destroy_op1)
alns.add_destroy_operator(destroy_op2)
alns.add_repair_operator(repair_op1, only_after=[destroy_op1])
alns.add_repair_operator(repair_op2)

In this example, repair_op1 will only be selected if destroy_op1 was selected, whereas repair_op2 can be used with both destroy_op1 and destroy_op2.

This has consequences for the WeightScheme.select_operators interface: it now takes an additional op_coupling argument, which is a binary 0/1 matrix indicating which operators can work together. In particular, element (i, j) is 1 if destroy operator i and repair operator j can work together, 0 otherwise. The WeightScheme has been updated to use this information to select the appropriate operator pairs.

Existing code should not be impacted, unless your code uses a custom weight scheme.

Version 3.0.1

Adds logging to the ALNS library, using the default Python logging module.

Version 3.0.0

Changelog:

Generalised stopping criteria (thanks @leonlan!) . Whereas previously only a maximum number of iterations was supported, you can now also use a maximum runtime.

Stopping criteria are available in alns.stop, and should be passed as the stop argument to ALNS.iterate. The stop argument replaces iterations.

Take the following example, using 10_000 iterations:

alns.iterate(..., iterations=10_000)

You would now write this as

from alns.stop import MaxIterations


alns.iterate(..., stop=MaxIterations(10_000))

In addition, alns.stop.MaxRuntime lets you set a maximum runtime (in seconds). You can also write your own stopping criterion, by inheriting from alns.stop.StoppingCriterion.

---

With the addition of stopping criteria, the old usage of the criteria module for acceptance criteria has become imprecise. This is why alns.criteria is now accessible as alns.accept. To simplify the module names further, alns.weight_schemes has been renamed to alns.weights.

---

ALNS now has a code style: black! The style is now consistent, and will be enforced through the CI pipeline.

---

Example notebooks and documentation have been updated to the new usage.

Version 2.1.2

Fixes a bug where key-word arguments passed to ALNS.iterate were not available in the on_best callback.

Version 2.1.1

Fix an off-by-one mistake in the number of runtimes to collect. We want to collect how long one iteration takes, but by accident collected the start time as well. This resulted in a mistaken value at index 0.

This release removes that mistake, such that the actual results are properly aligned.

Version 2.1.0

Collects iteration runtimes. These can be accessed as

res = alns.iterate(...)
runtimes = res.statistics.runtimes  # per iteration, in seconds

Version 2.0.2

Allow passing kwargs to ALNS.iterate. These kwargs are then passed to the operators. This is useful for larger heuristics, as this allows one to avoid explicit dependencies on global state (like a global data singleton).