MInCo: Mitigating Information Conflicts in Visual Model-based Reinforcement Learning

Introduction

This is a PyTorch implementation for the MInCo algorithm. MInCo is a visual model-based reinforcement learning method that mitigating information conflicts.

Instructions

1. Install Mujoco 2.1.0

2. Create an environment

   conda create -n minco python=3.8
   conda activate minco
   

3. Install dependencies

   pip install -r requirements.txt
   

Distracted DMC experiments

Use one of the following commands to train an agent on distracted Walker Walk. To train on other distracted DMC environments, replace walker-run with {domain}-{task}:

# MInCo
python experiments/train.py --algo minco --env_id dmc_distracted-walker-run --expr_name benchmark --seed 0 --a 8e-6 --prior_train_steps 5 --b 5 --c 0.015 --cross_inv_dynamics True

# RePo
python experiments/train.py --algo repo --env_id dmc_distracted-walker-run --expr_name benchmark --seed 0

# Dreamer
python experiments/train.py --algo dreamer --env_id dmc_distracted-walker-run --expr_name benchmark --seed 0

Visualization

Acknowledgements

We thank the RePo authors for their implementation of Pytorch version of Dreamer and Distracted DMC wrappers.