proger / curiosity

Unsupervised Exploration Agents

Agents with intrinsic rewards

Read more about curiosity in https://people.idsia.ch/~juergen/artificial-curiosity-since-1990.html

Environment setup.

python -m venv .venv
.venv/bin/pip install wandb celluloid safetensors
.venv/bin/pip install virtualenv
.venv/bin/pip install -r new_requirements.txt -I
.venv/bin/pip install -e ./gym-minigrid

Train RND agent. Reduce num_actors until everything starts to fit in VRAM.

ulimit -n 64000 # prevents RuntimeError: unable to open shared memory object </torch_13695_1684771047_984> in read-write mode: Too many open files (24)
env OMP_NUM_THREADS=1 .venv/bin/python -m src.algos.recurrent_rnd --intrinsic_reward_coef 0.1 --rnd_history 16 --rnd_autoregressive forward-target-difference --num_actors 30

Try the environment with manual control. Use ssh -X with XQuartz running on a Mac if needed.

.venv/bin/python gym-minigrid/manual_control.py # whole map view
.venv/bin/python gym-minigrid/manual_control.py --agent_view

Profile the environment. py-spy that write flamegraph SVGs on disk. This revealed that the original experience collection agents were spending most of their time running their policies on CPU. Another use let me know that actor workers are waiting a lot of time synchronous cuda transfer of useless tensors.

pip install py-spy
py-spy record --pid 3081044 -r 33 -s

Perform a rollout from a trained policy. This writes animation.mp4 on disk.

python -m src.algos.rnd exp/curiosity-20230901-121847/model.tar --fix_seed

Original README contents below:

RIDE: Rewarding Impact-Driven Exploration for Procedurally-Generated Environments

This is an implementation of the method proposed in

RIDE: Rewarding Impact-Driven Exploration for Procedurally-Generated Environments

by Roberta Raileanu and Tim Rocktäschel, published at ICLR 2020.

We propose a novel type of intrinsic reward which encourges the agent to take actions that result in significant changes to its representation of the environment state.

The code includes all the baselines and ablations used in the paper.

The code was also used to run the baselines in Learning with AMIGO: Adversarially Motivated Intrinsic Goals. See the associated repo for instructions on how to reproduce the results from that paper.

Citation

If you use this code in your own work, please cite our paper:

@inproceedings{
  Raileanu2020RIDE:,
  title={{RIDE: Rewarding Impact-Driven Exploration for Procedurally-Generated Environments}},
  author={Roberta Raileanu and Tim Rockt{\"{a}}schel},
  booktitle={International Conference on Learning Representations},
  year={2020},
  url={https://openreview.net/forum?id=rkg-TJBFPB}
}

Installation

# create a new conda environment
conda create -n ride python=3.7
conda activate ride 

# install dependencies
git clone [e-mail]:facebookresearch/impact-driven-exploration.git
cd impact-driven-exploration
pip install -r requirements.txt

# install MiniGrid
cd gym-minigrid
python setup.py install

Train RIDE on MiniGrid

cd impact-driven-exploration

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-MultiRoom-N7-S4-v0 --total_frames 30000000 --intrinsic_reward_coef 0.1 --entropy_cost 0.0005 --num_actors 3

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-MultiRoomNoisyTV-N7-S4-v0 --total_frames 30000000 --intrinsic_reward_coef 0.1 --entropy_cost 0.0005

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-MultiRoom-N7-S8-v0 --total_frames 30000000 --intrinsic_reward_coef 0.5 --entropy_cost 0.001

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-MultiRoom-N10-S4-v0 --total_frames 30000000 --intrinsic_reward_coef 0.1 --entropy_cost 0.0005

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-KeyCorridor-S3-R3-v0 --total_frames 30000000 --intrinsic_reward_coef 0.1 --entropy_cost 0.0005

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-ObstructedMaze-2Dlh-v0 --total_frames 100000000 --intrinsic_reward_coef 0.5 --entropy_cost 0.001

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-MultiRoom-N10-S10-v0 --total_frames 100000000 --intrinsic_reward_coef 0.5 --entropy_cost 0.001

OMP_NUM_THREADS=1 python main.py --model ride --env MiniGrid-MultiRoom-N12-S10-v0 --total_frames 100000000 --intrinsic_reward_coef 0.5 --entropy_cost 0.001

To train RIDE on the other MiniGrid environments used in our paper, replace the --env argument above with each of the following:

MiniGrid-MultiRoom-N7-S4-v0
MiniGrid-MultiRoomNoisyTV-N7-S4-v0
MiniGrid-MultiRoom-N7-S8-v0
MiniGrid-MultiRoom-N10-S4-v0
MiniGrid-MultiRoom-N10-S10-v0
MiniGrid-MultiRoom-N12-S10-v0
MiniGrid-ObstructedMaze-2Dlh-v0 
MiniGrid-KeyCorridorS3R3-v0

Make sure to use the best hyperparameters for each environment, as listed in the paper.

To run different seeds for a model, change the --run_id argument.

Overview of RIDE

Results on MiniGrid

Analysis of RIDE

Acknowledgements

Our vanilla RL algorithm is based on Torchbeast, which is an open source implementation of IMPALA.

License

This code is under the CC-BY-NC 4.0 (Attribution-NonCommercial 4.0 International) license.