info-structures/ais

This repository contains the code for RL for POMDPs through learning an Approximate Information State.

AIS based PORL

This repository contains the code for the PORL (partially observed reinforcement learning) experiments presented in the paper

J. Subramanian, A. Sinha, R. Seraj, and A. Mahajan, "Approximate Information State for Approximate Planning and Reinforcement Learning in Partially Observable Environments", 2020.

Three classes of experiments are presented (with their gym-environment-names)

• Low-dimensional environments

  • Tiger: Tiger-v0
  • Voicemail: Voicemail-v0
  • Cheese Maze: CheeseMaze-v0

• Moderate-dimensional environments

  • Rock Sample: RockSampling-v0
  • Drone Surveillance: DroneSurveillance-v0

• High-dimensional environments

  Various grid-world models from gym-minigrid (used in the BabyAI platform) including

  • Simple crossing: MiniGrid-SimpleCrossingS9N1-v0, MiniGrid-SimpleCrossingS9N2-v0, MiniGrid-SimpleCrossingS9N3-v0, MiniGrid-SimpleCrossingS11N5-v0
  • Lava crossing: MiniGrid-LavaCrossingS9N1-v0, MiniGrid-LavaCrossingS9N2-v0
  • Key corridor: MiniGrid-KeyCorridorS3R1-v0, MiniGrid-KeyCorridorS3R2-v0, MiniGrid-KeyCorridorS3R3-v0
  • Obstructed maze: MiniGrid-ObstructedMaze-1Dl-v0, MiniGrid-ObstructedMaze-1Dlh-v0
  • Misc MiniGrid-Empty-8x8-v0, MiniGrid-DoorKey-8x8-v0, MiniGrid-FourRooms-v0

Installation

To install all the dependencies of the code in a virtual environment, run the setup script:

bash bin/setup.sh

Usage

Fist activate the virtual environment using:

source python-vms/ais/bin/activate

To run AIS training algorithm for an environment, say Tiger-v0, run:

python src/main.py --env_name Tiger-v0 

This program accepts the following command line arguments:

Option	Description
--output_dir	The results are stored in this directory.
--env_name	The environment name (in open-ai gym format)
--eval_frequency	Number of batch iterations per evaluation step.
--N_eps_eval	Number of episodes to evaluate in an evaluation step.
--beta	Discount Factor
--lmbda	Trade-off between the next reward loss and next observation loss. It generally helps to keep this value low if the rewards of the environment are high.
policy_LR	Learning rate used by the ADAM optimizer for the policy.
ais_LR	Learning rate used by the ADAM optimizer for the ais.
batch_size	Number of samples used in a batch for every optimization step.
num_batches	Number of batches to train on.
AIS_state_size	Size of the vector used to represent the approximate information state.
--AIS_pred_ncomp	Number of components in the GMM for MiniGrid with the KL IPM.
--IPM	The IPM can be specified using this argument. In this code, MMD can be used to use the L2-norm squared form of the kernel based IPM. Or KL can be used to indirectly optimize for the Wasserstein IPM.
--seed	Random seed used.
--models_folder	Directory to save/load models

Reproducing results in the paper

The results presented in the paper can be obtained by running the following wrapper scripts:

• Low-dimensional environments

  • Tiger: sh bin/lowdim/tiger_MMD.sh and sh bin/lowdim/tiger_KL.sh
  • Voicemail: sh bin/lowdim/voicemail_MMD.sh and sh bin/lowdim/voicemail_KL.sh
  • Cheese Maze: sh bin/lowdim/cheesemaze_MMD.sh and sh bin/lowdim/cheesemaze_KL.sh

• Moderate-dimensional environments

  • Rock Sample: sh bin/moddim/rocksampling_MMD.sh and sh bin/moddim/rocksampling_KL.sh
  • Drone Surveillance: sh bin/moddim/dronesurveillance_MMD.sh and sh bin/moddim/dronesurveillance_KL.sh

• High-dimensional environments

  • sh bin/highdim/minigrid_MMD.sh and sh bin/highdim/minigrid_KL.sh

    This runs SimpleCrossingS9N1 environment. To run other environments, the name of the environment must be changed in bin/highdim/minigrid_*.sh files.

Citation

Please use the following citation to refer to the paper:

@misc{AIS,
      title={Approximate information state for approximate planning and reinforcement learning in partially observed systems}, 
      author={Jayakumar Subramanian and Amit Sinha and Raihan Seraj and Aditya Mahajan},
      year={2020},
      note={arXiv:2010.08843},
      url={https://arxiv.org/abs/2010.08843},

}