This repository contains code for the paper:

On the Critical Role of Conventions in Adaptive Human-AI Collaboration

"On the Critical Role of Conventions in Adaptive Human-AI Collaboration"
Andy Shih, Arjun Sawhney, Jovana Kondic, Stefano Ermon, Dorsa Sadigh
Proceedings of the 9th International Conference on Learning Representations (ICLR 2021)

@inproceedings{SSKESiclr21,
  author    = {Andy Shih and Arjun Sawhney and Jovana Kondic and Stefano Ermon and Dorsa Sadigh},
  title     = {On the Critical Role of Conventions in Adaptive Human-AI Collaboration},
  booktitle = {Proceedings of the 9th International Conference on Learning Representations (ICLR)},
  month     = {may},
  year      = {2021},
  keywords  = {conference}
}

Instructions

Install gym environment, hanabi environment, and stable baselines

pip install .

cd hanabi
pip install .
cd ../

cd stable-baselines3
pip install -e .
cd ../

Train partner agents

bash bashfiles/arms_train_selfplay.bash 1230 2
bash bashfiles/arms_train_selfplay.bash 1240 2

for ((i=1230;i<=1235;i++))
do
    bash bashfiles/blocks_train_selfplay.bash $i
done
for ((i=1240;i<=1245;i++))
do
    bash bashfiles/blocks_train_selfplay.bash $i
done
for ((i=1240;i<=1247;i++))
do
    bash bashfiles/hanabi_train_selfplay.bash $i
done

Run adaptation experiments. Choose from one of the settings:

• t=1: modular, regularization lambda=0.0
• t=2: modular, regularization lambda=0.3
• t=3: modular, regularization lambda=0.5
• t=4: baseline agg, aggregate gradients
• t=5: baseline agg, aggregate gradients, early stopping
• t=6: baseline modular, no main logits
• t=7: low-dim z + modular, regularization lambda=0.5

t=1
runid=100

bash bashfiles/arms_adapt_to_selfplay.bash $runid $t 2
bash bashfiles/arms_adapt_to_fixed.bash $runid $t 2
bash bashfiles/arms_human_adapt_to_fixed.bash $runid $t 2

bash bashfiles/blocks_adapt_to_selfplay.bash $runid $t
bash bashfiles/blocks_adapt_to_fixed.bash $runid $t

bash bashfiles/hanabi_adapt_to_selfplay.bash $runid $t