D-LM-MA (Distributed Low-Memory Matrix Adaptation evolution strategy)

The goal of this open-source repository is to provide all the source code and data involved in the paper Collective Learning of Low-Memory Matrix Adaptation for Large-Scale Black-Box Optimization, which has been submitted to PPSN-2022 (Under Review).

How-to-Run On Modern Cluster Computing

Configurations of the Python Programming Environment

It is suggested to use conda to create the virtual environment for Python.

Here, we only need to install two very popular libraries for numerical computing and distributed computing: NumPy and Ray.

$ conda create --prefix env_ppsn -y  # for virtual environment
$ conda activate env_ppsn/
$ conda install --prefix env_ppsn python=3.8.12 -y  # for Python
$ pip install numpy==1.21.5  # for numerical computing
$ pip install ray==1.9.1  # for distributed computing
$ pip install "ray[default]"

Settings of Environment Variables

According to the official suggestions from ray, we set the following environment variables, in order to avoid contention with multi-threaded libraries (NumPy here).

$ export OPENBLAS_NUM_THREADS=1
$ export MKL_NUM_THREADS=1
$ export OMP_NUM_THREADS=1
$ export NUMEXPR_NUM_THREADS=1

Note that the above settings should be applied in all computing nodes.

Construction of a Private Cluster Computing Platform

When CentOS is used, the firewall should be closed in advance for all nodes, as presented below.

$ sudo firewall-cmd --state
$ sudo systemctl stop firewalld
$ sudo systemctl disable firewalld
$ sudo systemctl mask --now firewalld
$ sudo firewall-cmd --state

We need to run one private Ray Cluster where the distributed algorithm is executed. There are several ways to construct it. For more guidelines about the Ray Cluster, see https://docs.ray.io/en/latest/cluster/quickstart.html. Here is the simplest way to construct it:

$ ray start --head  # run on the (single) master node, depending on your choice
# run on all slave nodes
$ ray start --address='[MASTER-ID:PORT]' --redis-password='[PASSWORD]'
# [MASTER-ID:PORT] and [PASSWORD] need to be replaced by your own settings
$ ray status  # run on the master node to check the status information of the Ray Cluster

How to Run Trials

Since typically the optimization process in high-dimensional cases needs a very long runtime, it is better to run these algorithms in the background (e.g., via nohup).

# run D-LM-MA (with 250 islands) independently for 7 times
$ nohup python run_trials.py -s=1 -e=7 -o=LMMAES -d=True -i=250 >DistributedLMMAES_1_7.out 2>&1 &
# run the baseline algorithm: serial MAES
$ nohup python run_trials.py -s=1 -e=7 -o=MAES >MAES_1_7.out 2>&1 &
# run the baseline algorithm: serial LMMAES
$ nohup python run_trials.py -s=1 -e=7 -o=LMMAES >LMMAES_1_7.out 2>&1 &

Note that the core code for D-LM-MA is available at https://github.com/Evolutionary-Intelligence/D-LM-MA/blob/main/pypoplib/distributed_es.py, while distributed_lmmaes.py is just its wrapper based on LM-MA-ES. In fact, other LSO versions of CMA-ES can also be used here with little modifications.