mork-experiments

mork-experiments

One Mork project per problem (CPH, CWP, MMDP, MDP). Each project re-implements the same algorithm logic as the corresponding project in ../jmh, but running on top of Mork's experimentation and instance-loading infrastructure instead of the hand-written experiment runner.

Implementation guide: ../../docs/mork-experiments.md documents how these projects were derived from the jmh "proto-framework" — first the general concept-by-concept conversion, then a section per problem.

Design principle: Mork infrastructure, jmh algorithms

Each project uses Mork for everything around the algorithm, and keeps the algorithm itself a faithful, hand-written port of jmh:

Concern	Provided by Mork	Ported from jmh (hand-written)
Instance model + loading	Instance, InstanceImporter	the parser and the data model
Solution + objective	Solution, Objective (FMode)	the representation, score & neighbourhood
Reproducible randomness	RandomManager, CollectionUtil	—
Time control	TimeLimitCalculator, TimeControl	the 10 s / 2 s budget per problem
Experiment runner, reports, dashboard	AbstractExperiment, Mork.start, Excel/JSON, web UI	which algorithms to compare
The algorithm	(not used yet)	multi-start local search / Scatter Search

Important — no reusable Mork algorithm components (yet). The algorithms deliberately do not use Mork's ready-made building blocks (SimpleAlgorithm, LocalSearchBestImprovement / LocalSearchFirstImprovement, GraspBuilder, MultiStartAlgorithm, ScatterSearchBuilder, VND, …). Instead each project contains a custom Algorithm that orchestrates hand-written constructive and improvement classes mirroring jmh. Migrating these to Mork's reusable components is a planned later step (that is where Move/Neighborhood will be exercised by Mork's own local searches; today a Move type exists only to declare the Objective).

Nodes are modelled as int indices with distances on the Instance (idiomatic Mork, as in the official TSP example) for CPH/CWP/MMDP; MDP keeps its MDPNode domain objects because its Scatter Search port relies on them.

Projects

Project	Problem	Sense	Algorithm (custom port)	Entry point
CPH	Capacitated p-hub	minimize	multi-start local search	es.urjc.etsii.grafo.cph.Main
CWP	Cutwidth	minimize	multi-start local search	es.urjc.etsii.grafo.cwp.Main
MMDP	MaxMin Diversity	maximize	multi-start local search	es.urjc.etsii.grafo.mmdp.Main
MDP	Maximum Diversity	maximize	Scatter Search	es.urjc.etsii.grafo.mdp.Main

Each project has the standard Mork layout: model/ (Instance, Importer, Solution, Move, Validator), constructives/ + improve/ (or algorithm/ for MDP), algorithms/ (the custom Algorithm), experiments/ (the AbstractExperiment + a TimeLimitCalculator), Main, and src/main/resources/{application.yml,serializers.yml}. Benchmark instances live under each project's instances/ folder.

Requirements

• JDK 25 (Mork 0.22).
• Maven (or the bundled toolchain). The projects depend on es.urjc.etsii.grafo:mork:0.22 via mork-parent:0.22 (resolved from Maven Central).

Build & run

cd CPH            # or CWP, MMDP, MDP
mvn clean package        # builds target/<Project>-1.0-SNAPSHOT.jar
java -jar target/<Project>-1.0-SNAPSHOT.jar

On start, Mork opens a live dashboard at http://localhost:8080/ and, when finished, writes an Excel report under results/. Runs are reproducible for a fixed solver.seed.

Useful overrides (Spring Boot style, --key=value):

• --instances.path.default=instances/<subset> — solve a different instance set or a single file.
• --solver.repetitions=N — repeat each (instance, algorithm). Default is 1 here, because each multi-start algorithm already performs 5000 internal restarts (Scatter Search runs to its time budget), reproducing one jmh run.
• -D<problem>.timeLimitMillis=MS (JVM property, before -jar) — the per-(instance, algorithm) time budget (default 10 s for CPH/CWP/MMDP, 2 s for MDP, matching jmh). Lower it for quick smoke runs, e.g. java -Dcph.timeLimitMillis=800 -jar ....

Notes:

• solver.benchmark is disabled in application.yml for a faster start; enable it to normalize time limits across machines.
• Algorithm short-names are limited by Mork to 30 characters.