metaheuristics

Fifteen metaheuristics written from scratch in C++ and compared on the same problem under the same evaluation budget.

Coursework for Metaheuristics, year 4 of the double degree in Computer Science and Business Administration, University of Granada (2025-26). Solo work.

Practices 1 to 3 solve the same problem: portfolio selection in the Markowitz sense. Given the historical returns of a market, choose the weight of each asset so the portfolio balances return against risk, with each weight inside a band (0.005 to 0.08 on IBEX 35) and the weights adding up to one.

Practice 4 changes the ground: the CEC'17 benchmark of continuous functions in 10, 30, 50 and 100 dimensions, which is what the field uses to compare optimisers.

The rule that shapes everything: a fixed budget of evaluations, the same for every algorithm, 50 independent runs each. An algorithm is not better because it finds a good solution, but because it finds it inside the budget.

Things worth naming:

• One shared framework. common/ holds the problem and solution abstractions, so every algorithm receives a Problem and returns a solution: swapping the metaheuristic does not touch the problem, and swapping the problem does not touch the metaheuristic.
• Experiments driven by a file. config.cfg fixes the seed, the evaluation budget, the number of runs and the market. Reproducing an experiment means editing that file, not recompiling.
• Every run is written out. Each practice leaves a CSV with one row per run, which is what feeds the box plots and the tables in the reports.
• Random search is always in the comparison. It is the floor: a metaheuristic that does not beat it is not doing any work.

src/practice-1..4/     the four practices, each with its own CMakeLists and config.cfg
  common/              problem and solution abstractions shared by the algorithms
  inc/  src/           one header and one file per metaheuristic
  *.csv                one row per run, the raw material of the reports
docs/reports/          the four LaTeX reports, with their figures and tables

• CMake 3.28 or later (practice 4 asks for 3.0) and a C++ compiler with C++17.
• Python 3 with numpy, pandas and matplotlib, only for the box plot and analysis scripts.

Each practice is an independent project:

cmake -S src/practice-1 -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j"$(nproc)"
cd src/practice-1 && ../../build/main        # reads config.cfg and the data next to it

Practice 4 ships its own script, which builds everything and runs a smoke test:

cd src/practice-4 && bash probar.sh          # about a minute, exit code 0 if it passes
bash probar.sh --completo                    # the full CEC'17 sweep, 30 to 40 minutes

Ranking on IBEX 35 after 50 runs of each algorithm, from the CSV in src/practice-3. Fitness is maximised, so higher is better:

The population-based algorithms with a local search on top take the top of the table, and the gap between them is smaller than the gap from any of them to the greedy solution. Random search sits last, which is the point of keeping it in the comparison.

The box plots per market and the statistical tests are in docs/reports/.

• The budget is the experiment. Once every algorithm gets the same number of evaluations, the comparison stops being about who converges lower and becomes about who spends the budget better. That reframing is most of what these four practices teach.
• Hybridising helps up to a point. The memetic variants beat their genetic parents, but ILS-ES-CHC, which stacks three ideas, lands mid-table: the extra machinery costs evaluations that the search then does not have.
• Limitations, kept as handed in:
  • Practice 3 does not compile as published. Its CMakeLists.txt includes common/, but that folder was never committed for this practice; practices 1, 2 and 4 do have it. The CI builds the other three and leaves this one out instead of patching it.
  • The large datasets are not in the repository. syp_100.csv and syp_500.csv (33 MB together) and the CEC'17 matrices for 30, 50 and 100 dimensions (39 MB) belong to the subject and to the benchmark suite, and are not mine to republish. What stays is ibex_35.csv and the CEC'17 files up to 20 dimensions, which is enough to run.
  • Identifiers and comments mix English and Spanish, and each practice repeats the framework instead of sharing one copy: they were handed in as separate deliverables.

Ismael Sallami Moreno. Released under the MIT licence (see LICENSE).