Currently, three benchmark suites are included:
Synthetic containing 24 noiseless functions, borrowed from coco:bbob with original paper.Noisy-Synthetic containing 30 noisy functions, borrowed from coco:bbob-noisy with original paper.Protein-Docking containing 280 problem instances, which simulate the application of protein docking as a 12-dimensional optimization problem, borrowed from LOIS with original paper.By setting the argument --problem to bbob, bbob-noisy or protein in command line to use the corresponding suite, for example:
python main.py --train --problem protein --train_agent MyAgent --train_optimizer MyOptimizer
For the usage of --train --train_agent --train_optimizer, see Training for more details.
Each test suites are regarded as a dataset, which is split into training set and test set in different proportions with respect to two difficulty levels:
easy training set accounts for 75% and test set accounts for 25%.difficult training set accounts for 25% and test set accounts for 75%.By setting the argument --difficulty to easy or difficult in command line to specify the difficulty level like the following command. Note that easy difficulty is used by default.
python main.py --train --problem bbob --difficulty difficult --train_agent MyAgent --train_optimizer MyOptimizer