johannes-dornheim/gTLO

generalized oulti-objective deep reinforcement learning algorithm

gTLO

generalized multi-objective deep reinforcement learning algorithm

Installation

(1) General:

• Manually install tensorflow-gpu up to v. 1.14 (e.g. conda install tensorflow-gpu=1.14 with python version <= 3.4.)

(2.a) For Deep-Sea Treasure Environment:

• Install ALE following https://github.com/garlicdevs/Fruit-API
• Install fruitAPI

(2.b) For Deep Drawing Environment:

• Prerequisite: abaqus (Tested on V. 6.14), Student version should be sufficient
• Install gym_fem following https://github.com/johannes-dornheim/Reinforce-FE

(3) General:

• install gTLO (pip install . in gTLO root folder)

Run presets / reproduce paper results

experiments are managed by the script agents/morl_agent.py and configured in ini files. To reproduce the results presented within the gTLO paper, the example configurations can be used as follows:

DST

• gTLO: python morl_agent.py --config ./preset_configs/DST_gTLO_250ksteps.ini
• outer-loop gTLQ: python morl_agent.py --config ./preset_configs/DST_gTLO_outerloop_25kSteps.ini
• gLinear: python morl_agent.py --config ./preset_configs/DST_gLinear_250kSteps.ini
• dTLQ (baseline agent): run study_starter.py from the FruitAPI fork https://github.com/johannes-dornheim/Fruit-API

Deep Drawing

• gTLO: python morl_agent.py --config ./preset_configs/DeepDrawing_gTLO.ini
• gLinear: python morl_agent.py --config ./preset_configs/DeepDrawing_gLinear.ini

Cite

@misc{https://doi.org/10.48550/arxiv.2204.04988,
  author = {Dornheim, Johannes},
  title = {gTLO: A Generalized and Non-linear Multi-Objective Deep Reinforcement Learning Approach},
  publisher = {arXiv},
  year = {2022},
  doi = {10.48550/ARXIV.2204.04988},
  url = {https://arxiv.org/abs/2204.04988},
  
}