/CS-GOMEA

Source code for paper Convolutional neural network surrogate-assisted GOMEA

Primary LanguageC

CS-GOMEA

Paper: Arkadiy Dushatskiy, Adriënne M. Mendrik, Tanja Alderliesten, and Peter A. N. Bosman. 2019. Convolutional neural network surrogate-assisted GOMEA. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO '19), Manuel López-Ibáñez (Ed.). ACM, New York, NY, USA, 753-761. DOI: https://doi.org/10.1145/3321707.3321760

Link to the paper: https://dl.acm.org/citation.cfm?id=3321760

Compilation

To compile CS-GOMEA: ./m_cs_gomea

To compile vanilla GOMEA: ./m_vanilla_gomea

Benchmark problems

Problems:

  1. Onemax
  2. Tight Trap4
  3. Loose Trap4
  4. NK Landscapes
  5. HIFF

Run algorithms

Several runs can be made, folders with results are created automatically:

  1. Convolutional Neural Net Surrogate-Assisted GOMEA (CS-GOMEA):

    python run_cs_gomea.py PROBLEM_NUMBER DIMENSIONALITY FOS_TYPE MAX_EVALUATIONS DELTA WARMUP_PERIOD FIRST_RUN N_RUNS DEVICE_ID TIME_LIMIT

  2. Vanilla GOMEA:

    python run_vanilla_gomea.py PROBLEM_NUMBER DIMENSIONALITY FOS_TYPE MAX_EVALUATIONS FIRST_RUN N_RUNS

  3. SMAC (https://github.com/automl/pysmac):

    python run_vanilla_gomea.py PROBLEM_NUMBER DIMENSIONALITY FIRST_RUN N_RUNS N_EVALUATIONS

  4. Hyperopt (implementation of Tree Parzen Estimator, https://github.com/hyperopt/hyperopt):

    python run_vanilla_gomea.py PROBLEM_NUMBER DIMENSIONALITY FIRST_RUN N_RUNS N_EVALUATIONS

Parameters description:

  1. PROBLEM_NUMBER - problem number chosen from above-mentioned problems
  2. DIMENSIONALITY - number of variables
  3. FOS_TYPE - FOS algorithm of GOMEA, 1 (the Linkage Tree) is recommended
  4. MAX_EVALUATIONS - maximum number of function evaluations (real ones) allowed
  5. DELTA - a parameter of CS-GOMEA determining how aggressive real evaluations are. The recommended value is 1.02
  6. WARMUP_PERIOD - the number of solutions in warm-up period of CS-GOMEA. This parameter is problem dependent, but it is suggested to generate at least 100 solutions to train the surrogate model.
  7. FIRST_RUN - the id of the first run of in a series of runs. While running experiments, folders with names P_S/R are created, where P, S, R are PROBLEM_NUMBER, DIMENSIONALITY and id of run respectively.
  8. N_RUNS - number of algorithm runs.
  9. DEVICE_ID - If there any CUDA devices, the device id. -1 means CPU usage. It is recommended to use a GPU for acceleration.
  10. TIME_LIMIT - algorithm time limit (in minutes)

Make plots

  1. Convergence plots:

    python convergence_plots.py PROBLEM_NUMBER FIRST_RUN N_RUNS SMAC_HYPEROPT

    Creating convergence plots. FIRST_RUN, N_RUNS indicate the folders with experiments to look in; SMAC_HYPEROPT indicates whether to include SMAC and Hyperopt runs on plots

  2. Scalability plots: python scalability_plots.py.

    Simply creating all scalability plots for all available problems instances