Analyzing Adaptive Parameter Landscapes in Parameter Adaptation Methods for Differential Evolution

This repository provides the C++ code used in the following paper:

Ryoji Tanabe, Analyzing Adaptive Parameter Landscapes in Parameter Adaptation Methods for Differential Evolution, ACM GECCO2020, pdf

This code is based on the old version of the COCO framework implemented in C (https://coco.gforge.inria.fr/):

Nikolaus Hansen, Anne Auger, Olaf Mersmann, Tea Tusar, Dimo Brockhoff, COCO: A Platform for Comparing Continuous Optimizers in a Black-Box Setting, CoRR abs/1603.08785 (2016), link

Requirements

I compiled this code with gcc version 7.5.0 on the Ubuntu 18.04 LTS. Python (2 or 3) is needed to run "src/createfolders.py". Ruby (>=2.5) is also needed to run "src/run_de.rb", which is an optional script to perform an experiment automatically.

Usage

Compile

$ git clone https://github.com/ryojitanabe/APL.git
$ cd APL
$ make

Simple example

$ python src/createfolders.py tmp 
$ ./de -coco_func_num 1 -prosize 20 -nfe 200000 -alg shade -pop 100 -strategy current_to_pbest_1 -p_rate 0.05 -cross bin -pop_rate -1 -arc_rate 1 -memory_size 10 -num_cheat_param_sampling 50 -out_dir tmp -run_type bbob-apl-analysis -in_file dummy.dat

The first Python command creates a directory "tmp" for the procedure of the COCO framework. The second command runs a DE algorithm on the 20-dimensional 3rd function instance of f1 (the Sphere function). More precisely, the parameters of the DE are as follows:

• Maximum number of function evaluations (nfe): 200000 (=10,000 * 20)
• Parameter adaptation method (alg): SHADE
• Population size (pop): 100
• Mutation strategy (strategy): current-to-pbest/1
• Pbest rate in current-to-pbest/1 (p_rate): 0.05
• Crossover method (cross): binomial crossover
• Population rate (pop_rate): -1
  • This sets the population size to "dim * pop_rate".
  • "pop\rate_=-1" means that the population size is determined by the argument "pop".
• Archive rate (arc_rate): 1
  • The actual archive size is pop * arc_rate. Thus, in this example, 100 * 1 = 100.
• Memory size in SHADE (memory_size): 10
• Memory size in SHADE (num_cheat_param_sampling): 50
  • This generates 50 * 50 parameter pairs of the scale factor F and the crossover rate C.
• Output directory (out_dir): tmp
• Run type (run_type): bbob-apl-analysis
  • This runs DE to obtain adaptive parameter landscapes.
  • When running DE for benchmarking, it should be set to "bbob".
• Problem instance (in_file): dummy.dat
  • This determines which function instance should be used for the experiment.

This command outputs many dat files in the "tmp" directory such as "parameters_bbob-f1_d20_9th_ranked_ind_11000th_evals_3th_run.dat". This file contains 50 * 50 parameter pairs (F and C) and their corresponding fitness improvement values (the G1 value) of the 9th ranked individual in the population at 1,000 function evaluations on the 20-dimensional 3rd function instance of f1.

Similarly, the following is the command to run a DE with the parameter adaptation method in jDE:

$ python src/createfolders.py tmp 
$ ./de -coco_func_num 1 -prosize 20 -nfe 200000 -alg jde -pop 100 -strategy current_to_pbest_1 -p_rate 0.05 -cross bin -pop_rate -1 -arc_rate 1 -tau_sf 0.1 -tau_cr 0.1 -num_cheat_param_sampling 50 -out_dir tmp -run_type bbob-apl-analysis -in_file dummy.dat

The following is also the command to run a DE with the parameter adaptation method in JADE:

$ python src/createfolders.py tmp 
$ ./de -coco_func_num 1 -prosize 20 -nfe 200000 -alg jade -pop 100 -strategy current_to_pbest_1 -p_rate 0.05 -cross bin -pop_rate -1 -arc_rate 1 -jade_learning_rate 0.1 -num_cheat_param_sampling 50 -out_dir tmp -run_type bbob-apl-analysis -in_file dummy.dat

Reproduce results presented in the GECCO paper

The Ruby script "run_de.rb" automatically performs all experiments to reproduce all results presented in the GECCO paper:

$ ruby run_de.rb

This command outputs results in the "results" directory. The Ruby script refers to a file in "median_run_ids", which provides an index number of a single run with a median performance among 15 runs for each function.