This codebase contains the implementation for the paper PD-MORL: Preference-Driven Multi-Objective Reinforcement Learning Algorithm (ICLR 2023).
In this paper, we propose a novel preference-driven multi-objective reinforcement learning algorithm using a single policy network that covers the entire preference space in a given domain.
- Operating System: tested on Ubuntu 18.04.
- Python Version: >= 3.8.11.
- PyTorch Version: >= 1.8.1.
- MuJoCo : install mujoco and mujoco-py of version 2.1 by following the instructions in mujoco-py.
You can either install the dependencies in a conda virtual env or manually.
For conda virtual env installation, simply create a virtual env named pdmorl by:
conda env create -f environment.yml
If you prefer to do it manually, you could just simply open the environment.yml in the editor to install packages using pip.
You still need to satisfy the Prerequisites. For example, to install pytorch 1.8.1 with the cuda 11.1 toolkit, simply run the following command
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
All training and evaluation-related codes are in the folder PD-MORL. We also provide some common utilities and environments in the lib folder. We follow a similar structure to PTAN in our codebase for readibility and simplicity.
We evaluate PD-MORL's performance on two commonly used discrete MORL benchmarks: Deep Sea Treasure and Fruit Tree Navigation. For these benchmarks, we use the implementation of https://github.com/RunzheYang/MORL. We also evaluate its performance on multi-objective continuous control tasks such as MO-Walker2d-v2, MO-HalfCheetah-v2, MO-Ant-v2, MO-Swimmer-v2, and MO-Hopper-v2. For these benchmarks, we use the implementation of https://github.com/mit-gfx/PGMORL.
The code for the environments can be found in lib/utilities/morl/moenvs folder. You should register these environments as a gym environment by:
cd ../lib/utilities/morl/MOEnvs
pip install -e .
Additionally, if needed, you may comment out the following line 21 in this gym script home/[user_id]/.conda/envs/[env_name]/lib/python[version]/site-packages/gym/wrappers/time_limit.py:
# info["TimeLimit.truncated"] = not done
You can find the pretrained models in EvaluationNetworks folder. Using these models, you can run
python eval_benchmarks_MO_TD3_HER.py --benchmark_name 'walker' --model_ext 'actor' --eval_episodes 1
for continuous controls tasks. In this code, change the benchmark name and the model name accordingly. You can run either of these commands for discrete benchmarks
python test_DeepSeaTreasure_MO_DDQN_HER.py --model_ext 950370
python test_FruitTreeNavigation_MO_DDQN_HER.py --tree_depth 6 --model_ext 312050
In this code, change the model name accordingly.
You can also find precomputed Pareto front solutions in PD-MORL as objs_pdmorl_[problem name].
We provide a configuration of hyperparameters for each benchmark in lib/utilities/settings.py. You can modify these hyperparameters according to your needs.
We provide a training script in the PD-MORL folder for each benchmark. For instance, for the MO-Ant-v2 problem, you can run the following command to start training:
cd PD-MORL
python train_Walker2d_MO_TD3_HER.py
By default, the models are stored in Exps/[problem_algorithm name] during training. The Pareto front plots are stored in Figures/[problem name] for continuous benchmarks.
We run all our experiments on a local server, including Intel Xeon Gold 6242R @4.1GHz and 385 GB memory. We do not use any GPU in our implementation.
If you find our paper or code is useful, please consider citing:
@inproceedings{basaklar2022pd,
title={PD-MORL: Preference-Driven Multi-Objective Reinforcement Learning Algorithm},
author={Basaklar, Toygun and Gumussoy, Suat and Ogras, Umit},
booktitle={The Eleventh International Conference on Learning Representations},
year={2022}
}