Training, evaluation and visualization of how deep reinforcement learning policies divide the state space from Understanding the Evolution of Linear Regions in Deep Reinforcement Learning.

This codebase was tested with Python3.7, Pytorch1.12.1 and CUDA10.2 (when required). Clone the repository and navigate to the root directory of the repository:

cd deep_rl_regions

Install the required packages using:

pip install -r requirements.txt

Training logs are created using Weights & Biases (wandb) and evaluation metrics are also logged to wandb. Sign up on wandb using desired USERNAME, and create a project named deep_rl_regions. The chosen username and project name will be later used when running some of the scripts.

If you prefer not to use wandb, some effort is altering the scripts.

python train.py --device=cpu --env=HalfCheetah-v2 --total_timesteps=100000 --nsteps=1000 --eval_every=1 --save_every=1 --policy_dims=c16,16 --value_dims=c32,32 --save_dir=checkpoints --entity=USERNAME --project_name=deep_rl_regions

Environment can simply be changed to any Mujoco environment.

To evaluate a single training run with wandb run id of RUNID, run:

python evaluate.py --run_id=RUNID --entity=USERNAME --project_name=deep_rl_regions

To evaluate all runs on a wandb training sweep with SWEEPID, run:

python evaluate.py --sweep_id=SWEEPID --entity=USERNAME --project_name=deep_rl_regions

To compute metrics over random lines and random trajectories as described in section 5.2 of the paper, run the following script before running evaluate.py:

python randomized_metric_helper.py --sweep_id=SWEEPID

then run evaluate.py with options --evaluate_random_trajectories and --evaluate_random_lines.

To visualize lineare regions in high-dimensional state spaces, we define a 2-dimensional plane using 3 randomly sampled points in the state space, and project the linear regions onto this plane and visualize the projected linear regions. To do this, run:

python visualize_hd.py --env=HalfCheetah-v2 --policy_path=path_to_saved_policy_checkpoint --stats_path=path_to_saved_stats_checkpoint --save_path=path_to_saved_plot_dir

To visualize 2-dimensional state spaces, run:

python visualize_hd.py --env=envs:Car1DEnv-v1 --policy_path=path_to_saved_policy_checkpoint --stats_path=path_to_saved_stats_checkpoint --save_path=path_to_saved_plot_dir

@article{cohan2022understanding,
  title={Understanding the Evolution of Linear Regions in Deep Reinforcement Learning},
  author={Cohan, Setareh and Kim, Nam Hee and Rolnick, David and van de Panne, Michiel},
  journal={Advances in Neural Information Processing Systems},
  volume={35},
  pages={10891--10903},
  year={2022}
}