This repository contains the code for paper (...) ICML 2020.

Clone the repo and update the submodules:

git clone
cd repo
git submodule update --init --recursive

Install the dependancies:

conda env create -f env.yaml
conda activate gacem

To re-produce results icml_command.sh contains all the commands that were used to train and plot the results. Just find which command you want to run and un-comment the relative parts. For example if you want to run all the algorithms for function "Stybnlinski-2D" and get the plots uncomment the following lines:

# ---------- styblinski 2D
# CPU compatible
./run.sh scripts/run_experiments.py spec_files/icml_paper/styblinski_dim/styblinski_2d_cem_dim_study.yaml -ns 5
./run.sh scripts/run_experiments.py spec_files/icml_paper/styblinski_dim/styblinski_2d_cem_sg_fixed_sigma_dim_study.yaml -ns 5
./run.sh scripts/run_experiments.py spec_files/icml_paper/styblinski_dim/styblinski_2d_cempp_gauss_dim_study.yaml -ns 5
./run.sh scripts/run_experiments.py spec_files/icml_paper/styblinski_dim/styblinski_2d_cempp_kde_dim_study.yaml -ns 5
./run.sh scripts/run_experiments.py spec_files/icml_paper/styblinski_dim/styblinski_2d_gacem_off_fixed_sigma_dim_study.yaml -ns 5
./run.sh scripts/run_experiments.py spec_files/icml_paper/styblinski_dim/styblinski_2d_gacem_onp_fixed_sigma_dim_study.yaml -ns 5

# plot figures
./run.sh scripts/plot_experiments.py \
data/styblinski_2d_cem_gauss_optim_on_co0p4_ns25 \
data/styblinski_2d_cem_gauss_optim_on_co0p4_ns25_sig10 \
data/styblinski_2d_cem_gauss_optim_off_co0p4_ns25 \
data/styblinski_2d_cem_kde_optim_off_co0p4_ns25 \
data/styblinski_2d_gacem_ns5_lay3x100_e10_b10_nr40_csp_off_fixed_sigma \
data/styblinski_2d_gacem_ns25_lay3x100_e10_b10_nr40_csp_onp_fixed_sigma \
--save-path experiments_results/styblinski_2d \
-l cem_adaptive_variance cem_fixed_variance cem++sg cem++kde gacem_off gacem_onp \
-g 10

The following algorithms are implemented:

• Continuous domain adaptive sampling with Autoregressive policies
• Cross Entropy Method (Single Gaussian (SG), KDE (Kernel Desnsity Estimation), with adaptive co-variance, with fixed variance)

CEM has different variants: Optimization vs. Constraint satisfaction, with on-policy or off-policy

Modules reside in scr/optnet, some of the important ones are the following:

• alg/autoregressive/cont_autoreg_optim.py: gacem code
• alg/cross_entropy/cem_optim.py: cem code
• alg/utils/weight_compute.py: weight computation code
• benchmarks/functions.py: test functions
• data/buffer.py: Replay Buffer data structure
• models/made.py: MADE model
• torch: torch compatible distribution functions
• viz: visualization functions

top level scripts are located in ./scripts

The setting for each experiment is passed to top level scripts using yaml files: It basically contains the name of the algorithm class and the parameters passed to it, plus some info about the root directory for saving results: As an example look at spec_files/autoregressive/synt.yaml

to run:

cd autoreg_ckt
./run.sh scripts/run_alg.py --help
./run.sh scripts/run_alg.py spec_files/cem/synt.yaml
./run.sh scripts/run_alg.py spec_files/autoregressive/synt.yaml

to load the results and resume if possible (maybe the number of iterations was not sufficient and you want to load the new yaml and then resume):

./run.sh scripts/run_alg.py data/path_to_spec_yaml.yaml --load

to plot tsne or pca:

./run.sh scripts/compare_solutions.py --help

to animate 2D

Results are stored in data folder according to the name, suffix, prefix specified in the yaml and the date and time of running the algorithm.