PerSim: Data-Efficient Offline Reinforcement Learning with Heterogeneous Agents via Personalized Simulators

This is the code accompaying the paper submission PerSim: Data-Efficient Offline Reinforcement Learning with Heterogeneous Agents via Personalized Simulators"

Requirements

• python >3.6
• Mujoco-py and its prerequisites.
• python packages in requirements.txt

Datasets

We provide the offline datasets we performed the experiments on. The datasets can be downloaded via running data.sh through:

`bash data.sh`

Running PerSim

To run PerSim, run the following script:

`python3 run.py --env {env} --dataname {dataname} --r {rank}`

Choose env from {mountainCar, cartPole, halfCheetah}, and dataname from the available datasets in the datasets directory. e.g., cartPole_pure_0.0_0. Best values for r is 3,5,15 for mountainCar, cartPole, and halfCheetah respectively.