Lexicographic Multi-Objective Reinforcement Learning
Installation
To view installation instructions, go to INSTALLATION.md
Running Experiments
Default
To test your installation, run:
python -m src.run_batches humble_batch
Some individual experiments may be run using:
python -m src.train --agent_name=_ --env_name=_ --num_episodes=_
Running Batches
To simplify training and running experiments, we use a "batch" system.
The class TrainingParameters specifies all of the information needed to train a single agent on a particular environment.
The file batch_definitions contains "batches": lists of instances of TrainingParameter used to specify a sequence of training runs.
For example, the following batch would compare DQN and AC in the CartSafe environment:
"DQN_vs_AC": [
TrainingParameters( agent_name="DQN", env_name="CartSafe"),
TrainingParameters( agent_name="AC", env_name="CartSafe"),
],You can run these predefined batches using the run_batches.py file:
python -m src.run_batches DQN_vs_AC
Project structure
As above, run_batches.py and batch_definitions.py define meta-level code for managing experiments.
The file train.py run individual experiments and controls the state-agent interaction loop.
Individual agents are defined in the agents directory and environments in the envs.py file.
The graphing directory contains code for generating the graphs included in the paper.