Fuse2Control (F2C)

Implementation of our work 'Information-Theoretic State Space Model for Multi-View Reinforcement Learning' accepted at ICML 2023.

1. Prerequisites

Install following libraries and packages in a conda environment:

conda create -n fuse2control python=3.7.5
conda activate fuse2control
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch
pip3 install -r requirements.txt
pip install Box2D sklearn wget

2. Train & evaluate MVSSM and baseline algorithms

We provide algorithms included in the paper, e.g. MVSSM (ours), MVTCAE, SLAC, CMC or Vanilla-RL. Experiments can be started by running main.py as follows:

python main.py --method=$METHOD --missing_view_num=$MISSING_VIEW_NUM --seed=$SEED --use_collected_data=$USE_COLLECTED_DATA

• To choose among running the MVSSM, MVTCAE, SLAC, CMC and Vanilla-RL, you need to set the value of variable $METHOD to MVSSM, MVTCAE, SLAC, CMC or Vanilla-RL respectively.
• To choose the number of missing views, you need to change the value of variable $MISSING_VIEW_NUM to 0, 1, 2, 3 or 4 respectively.
• To choose the random seed, you need to change the value of variable $SEED to 0, 1, 2, 3 or 4 respectively.
• To collect dataset on your own (instead of using pre-collected dataset), you need to change the value of variable $USE_COLLECTED_DATA to False.

(Example)

python main.py --method=MVSSM --missing_view_num=0 --seed=0 --use_collected_data=True

3. Results

Notes

We used 40 CPU instances (n1-highcpu-32) from Google Cloud Platform(GCP).

All the numbers are averaged over 5 different seeds (0~4).

We appreciate PPO-PyTorch which our implementation is heavily based on.