Permanent Dropout: Stochastic Neuron Pruning for Accelerating Deep Reinforcement Learning

Installation

We recommend using Miniconda to create your virtual env.

conda create -n dropout_env python=3.11
conda activate dropout_env
git clone https://github.com/hail-mary/permanent-dropout.git
cd permanent-dropout
pip install gymnasium[mujoco] stable-baselines3 pyyaml

Training

python main.py

# option: specify the log directory by adding `--logdir` flag.
python main.py --logdir logs

Evaluation

python main.py --eval [PATH_TO_CHECKPOINTS]

# option: recording requires ->> pip install "gymnasium[other]"
python main.py --eval [PATH_TO_CHECKPOINTS] --record

Plot results

--plot can also accept multiple json files for comparison.

python main.py --plot .\logs\history.json

References

• Simultaneous Optimization of Discrete and Continuous Parameters Defining a Robot Controller
• Averaging Weights Leads to Wider Optima and Better Generalization
• Improving Stability in Deep Reinforcement Learning with Weight Averaging