Real-time Neural Networks

Implementation of Real-time Neural Networks for agent-state construction with a focus on constructive approaches

Requirements

In order to run this project, you'll need the following things installed:

• GCC (version 9.3.0 and above)
• make
• A C++ compiler (C++17 and above)
• MariaDB (and a C++ connector for MariaDB found here)

For the tests, you'll need Python installed together with pytorch and autograd.

To compile this project locally, you'll have to link your MariaDB C++ connector locally, as demonstrated in CMakeLists.txt. You'll also need to uncomment essentially everything under the comment "For running locally".

Instructions for running experiments in Python

• Install packages pip install -r requirements.txt
• Install pybind11 and adjust the pybind directory in CMakeListsPy.txt (recommended to install it as subdir in this project)
• Use CMakeListsPy.txt to compile

Mountain Car control experiment

• Train using: Python train_gym.py --env-max-step-per-episode 1000 -t control --tilecoding 1

Atari prediction experiments

• From the project's root directory, use git clone --recursive https://github.com/DLR-RM/rl-baselines3-zoo to get the pretrained expert agents
• Train using: python train_gym.py --net imprintingAtari --imprinting-mode random --env PongNoFrameskip-v4 --binning 1 --gamma 0.5 --meta-step-size 0.01