mltests: A Roff repository from knc-neural-calculus

mltests

About

Hi all,

This is a project with the loose goal of finding interesting ways to generate intelligent agents through trial and error. There are currently distinct Python and C++ architectures, each with a default basic spiking neural network and a experimental frequency-domain based spiking neural network. This document includes instructions on how to work with each architecture. The docker development environment is intended for use with the python architecture. Here is a showcase video, showing the behavior of some of the topologies stored in this repository: https://www.youtube.com/watch?v=zFZZjQkz7RA

Building dev Docker image

To set up the mltests dev environment with Docker, run the following commands in bash:

cd mltests
docker build -t mltests-dev .

vscodium and vim are included by default, feel free to add your own editor to the Dockerfile to suit your own preferences. In GUI mode, you can use vscodium! Git is included by default, so you can use git directly with your own fork of the mltests project from inside the Docker container.

Running dev Docker image with GUI(Recommended)

xhost +SI:localuser:root
docker run  -e DISPLAY=$DISPLAY \
            -v /tmp/.X11-unix:/tmp/.X11-unix:rw \
            --ipc=host \
            --user 0:0 \
            --cap-drop=ALL \
            --security-opt=no-new-privileges \
	    -it \
            mltests-dev

Adapted from https://github.com/mviereck/x11docker/wiki/Short-setups-to-provide-X-display-to-container.

Please note that this breaks container isolation!

Running dev Docker image without GUI:

docker run -it --user 0:0  mltests-dev

Revisting Existing Docker session

To determine the name of your session, get a list of docker processes with

docker ps -a

If the Docker session you are starting needs graphical windows, you will need the following line:

xhost +SI:localuser:root

Then to start the session (with changes saved) type:


docker start -a -i <session name>

This is an important idea to get the hang of, since building from scratch pulls from the online git repo.

Python architecture

cd python_neuro_ev
python3 ./main.py <mode> <environment> <learning_algorithm> <file_name>

modes:

train

runs neuro-evolution from scratch, saving generated toplogy to ./log.txt

improve

runs neuro-evolution, starting from an existing topology, and saves to ./log.txt as higher-scoring topologies are found

analyze

provides a visualization of the performance of the toplogy specified by the provided file on the microworld task.

environments:

cartpole chopper berzerk xor biped pshoppe sologrid

C++ architecture

enter the following commands in bash, from the project directory

make
chmod +x ./test
cd ./microworld1
make
chmod +x ./microworld1
cd ../microworld2
make
chmod +x ./microworld2
You're good to go!

microworlds 1 and 2 can be run from the command line in three modes

./microworld1 train