The project solves Bipedal-Walker-v3 using deep neural networks trained by genetic algorithm. We used simplified version of Deep Neuroevolution: Genetic Algorithms Are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning. Our best model reached average return of ~310 over 100 episodes.
- Clone the repository and change current directory
git clone https://github.com/standa42/bipedal-walker-deep-ga
cd "bipedal-walker-deep-ga"- Create virtual environment
- you can also reuse an existing one but this tutorial acts as you would create a new one
/usr/bin/python3 -m venv "venv"- Install pip requirements
venv/bin/python3 -m pip install -r "requirements.txt"Training is simple using following script:
venv/bin/python3 train.pyDefault parameters are already set for the best reached model. Logs are stored into logs/train{TIMESTAMP}_{UUID4_CODE} directory.
Evaluation of model is done using following script:
venv/bin/python3 evaluate.pyEvaluation can be visualized using render_each parameter.