amiithinks / mon_mdp_aamas24 Goto Github PK

Source code of the paper Monitored Markov Decision Processes.

To install and use our environments, run

pip install -r requirements.txt
cd src/gym-monitor
pip install -e .

To test and render an environment, run python and then

import gymnasium
env = gymnasium.make("Gym-Monitor/Gridworld-Medium-3x3-v0", render_mode="human")
env.reset()
env.step(1)
env.render()

Red tiles denote negative rewards, green tile is the goal, blue circle is the agent, the arrow means that the agent moved down.

We use Hydra to configure our experiments.
Hyperparameters and other settings are defined in YAML files in the configs/ folder.
Most of the configuration is self-explanatory. Some keys you may need to change are the following:

• WandB settings and Hydra log directories in configs/default.yaml,
• Folder experiment.datadir in configs/default.yaml (where npy data is saved),
• Folder experiment.debugdir in configs/default.yaml (where agent pics are saved),
• Steps, learning rate, epsilon decay, and other training parameters in configs/experiment/.

To try the Oracle algorithm on the Penalty MonMDP with the default configuration, and save some debug data, run

python main.py monitor=binary_stateless agent.critic.strategy=oracle experiment.debugdir=debug

This will save pics to easily visualize the Q-function and the greedy policy.
Everything will be saved in debug/, in subfolders depending on the environment and the monitor IDs.

In the policy pic (left), arrows denote the action executed in each tile.

• Orange arrows do not ask for monitoring,
• Empty white arrows do ask for monitoring,
• Orange arrows with white borders can do both randomly.

The policy is random in the goal tile because that is a terminal state.

In the Q-function pic (right), the heatmaps denote the value of the 4D Q-table, where Q[i,j,k,l] is the Q-value of environment state i, monitor state j, environment action k, monitor action l.

For a sweep over multiple jobs in parallel with Joblib, run

python main.py -m hydra/launcher=joblib hydra/sweeper=medium_det

Custom sweeps are defined in configs/hydra/sweeper/.
You can further customize a sweep via command line. For example,

python main.py -m hydra/launcher=joblib hydra/sweeper=medium_det experiment.rng_seed="range(0, 10)" monitor=limited_time hydra.launcher.verbose=1000

Configs in configs/hydra/sweeper/ hide the training progress bar of the agent, so we suggest to pass hydra.launcher.verbose=1000 to show the progress of the sweep.

If you have access to a SLURM-based cluster, you can submit multiple jobs, each running a chunk of the sweep with Joblib. Refer to submitit_jobs.py for an example.

Experiments will save the expected discounted return of the ε-greedy (training) and greedy (testing) policies in npy files (default dir is data/).
If you want to zip and copy only the data needed for plotting, run

find data -type f -name "*test*.npy" -print0 | tar -czvf data.tar.gz --null -T -

To plot expected return curves, use plot_curves.py. This script takes two arguments:

• -c is the config file that defines where to save plots, axes limits, axes ticks, what algorithms to show, and so on. Default configs are located in configs/plots/.
• -f is the folder where data from the sweep is located.

For example, running

python plot_curves.py -c configs/plots/deterministic_appendix.py -f data/iGym-Monitor/

Will generate many plots like these two, and save them in data/iGym-Monitor/deterministic_appendix.

Finally, python plot_legend.py will generate a separate pic with only the legend.