Metaheuristics Weight Finding Project

Overview

This project implements and evaluates various metaheuristic algorithms for weight finding in machine learning models, particularly for the Feature Weight Learning (FWL) problem. It includes a range of algorithms from simple local search methods to more complex population-based approaches, designed to optimize feature selection and weight assignment in classification tasks. The project uses 5-fold cross-validation (5CV) for analysis and comparison, focusing on weight generation for the 1-Nearest Neighbor (1NN) classification algorithm.

NOTE: The report is written in Spanish.

Project Structure

.
├── bin/                  # Compiled binaries (contains metaheuristics executable)
├── data/                 # ARFF data files and FWL tables for 2023-24
├── docs/                 # Documentation
│   ├── html/             # Doxygen-generated HTML docs
│   └── latex/            # Doxygen-generated LaTeX docs
├── include/              # Header files
│   ├── algorithms/       # Algorithm-specific headers
│   └── utils/            # Utility headers
├── obj/                  # Compiled object files
├── outputs/              # Execution results, generated solutions, and training fitness logs
├── scripts/              # Utility scripts (including run.sh and generate_training_fitness_table.py)
├── src/                  # Source files
│   ├── algorithms/       # Algorithm implementations
│   └── utils/            # Utility implementations
├── tests/                # Unit tests and mocks
└── Makefile              # Project compilation instructions

Datasets

The project evaluates the algorithms using three datasets:

• breast-cancer
• ecoli
• parkinsons

Algorithms Implemented

Local Search Methods

1. Local Search
2. Best Local Search (BLS)
3. Simulated Annealing (SA)
4. Iterated Local Search (ILS)
5. ILS with SA (OILS-SA)

Population-based Methods

6. Genetic Algorithm (GA)
   • AGE-CA: Steady-state GA with arithmetic crossover
   • AGE-BLX: Steady-state GA with BLX-α crossover
   • AGG-CA: Generational GA with arithmetic crossover
   • AGG-BLX: Generational GA with BLX-α crossover
7. Memetic Algorithm (MA)
8. MA with Restart
9. BMB (Baldwinian MA)

Other Methods

10. 1-Nearest Neighbor (1NN)
11. Relief
12. Random Search

Installation

1. Clone the repository:

   git clone https://github.com/pab1s/attribute-weight-metaheuristics.git
   cd mattribute-weight-metaheuristics
   

2. Compile the project:

   make
   

Usage

Run the compiled binary with appropriate arguments:

./scripts/run.sh [options]

Execution Parameters

The program accepts the following parameters:

• --algorithm=ALGORITHM_NAME: Specifies the algorithm to run (e.g., local-search, age-ca, memetic)
• --dataset=DATASET_NAME: Specifies the dataset to use (breast-cancer, ecoli, or parkinsons)
• --seed=SEED_VALUE: Sets the random seed
• --log: Enables logging of execution results

Additional algorithm-specific parameters can be specified. For example:

• Local Search:

  • --maxEvaluations: Maximum number of evaluations (default: 15000)
  • --maxNeighbors: Maximum number of neighbors to explore per iteration (default: 20)
  • --variance: Variance for the normal distribution in neighbor generation (default: 0.3)

• Genetic Algorithms:

  • --maxEvaluations: Maximum number of objective function evaluations (default: 15000)
  • --populationSize: Population size (default: 50)
  • --crossoverRate: Crossover probability (default: 1.0)
  • --mutationRate: Mutation probability (default: 0.08)

For specific usage instructions and parameters for each algorithm, please refer to the documentation in the docs/ directory.

Scripts

The scripts/ directory contains useful scripts:

• run.sh: Main execution script
• generate_training_fitness_table.py: Generates performance graphs

To run the main experiment script:

./scripts/run.sh

Documentation

This project uses Doxygen for documentation. To generate the documentation:

1. Ensure you have Doxygen installed.
2. Run:

   doxygen config.doxy
   

3. Access the generated documentation in docs/html/index.html or docs/latex/refman.pdf.

Contact

Pablo Olivares Martínez - [email protected]