Training a Resilient Q-Network against Observational Interference, Causal Inference Q-Networks

Last update: Nov 21, 2022

Overview

Obs-Causal-Q-Network

AAAI 2022 - Training a Resilient Q-Network against Observational Interference

Environment Setup

option 1 (from conda .yml under conda 10.2 and python 3.6)

conda env create -f obs-causal-q-conda.yml

option 2 (from a clean python 3.6 and please follow the setup of UnityAgent 3D environment for Banana Navigator )

pip install torch torchvision torchaudio
pip install dowhy
pip install gym

1. Example of Training Causal Inference Q-Network (CIQ) on Cartpole

Run Causal Inference Q-Network Training (--network 1 for Treatment Inference Q-network)

python 0-cartpole-main.py --network 1

Causal Inference Q-Network Architecture

Output Logs

observation space: Box(4,)
action space: Discrete(2)
Timing Atk Ratio: 10%
Using CEQNetwork_1. Number of Params: 41872
 Interference Type: 1  Use baseline:  0 use CGM:  1
With:  10.42 % timing attack
Episode 0   Score: 48.00, Average Score: 48.00, Loss: 1.71
With:  0.0 % timing attack
Episode 20   Score: 15.00, Average Score: 18.71, Loss: 30.56
With:  3.57 % timing attack
Episode 40   Score: 28.00, Average Score: 19.83, Loss: 36.36
With:  8.5 % timing attack
Episode 60   Score: 200.00, Average Score: 43.65, Loss: 263.29
With:  9.0 % timing attack
Episode 80   Score: 200.00, Average Score: 103.53, Loss: 116.35
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 193.4
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 164.2
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 147.8
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 193.4
With:  9.5 % timing attack
Episode 100   Score: 200.00, Average Score: 163.20, Loss: 77.38
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 198.4
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 197.8
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 197.6
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 198.6
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 199.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 186.8
Using CEQNetwork_1. Number of Params: 41872
### Evaluation Phase & Report DQNs Test Score : 200.0

Environment solved in 114 episodes!     Average Score: 195.55
Environment solved in 114 episodes!     Average Score: 195.55 +- 25.07
############# Basic Evaluate #############
Using CEQNetwork_1. Number of Params: 41872
Evaluate Score : 200.0
############# Noise Evaluate #############
Using CEQNetwork_1. Number of Params: 41872
Robust Score : 200.0

2. Example of Training a "Variational" Causal Inference Q-Network on Unity 3D Banana Navigator

Run Variational Causal Inference Q-Networks (VCIQs) Training (--network 3 for Causal Variational Inference)

python 1-banana-navigator-main.py --network 3

Variational Causal Inference Q-Network Architecture

Output Logs

'Academy' started successfully!
Unity Academy name: Academy
        Number of Brains: 1
        Number of External Brains : 1
        Lesson number : 0
        Reset Parameters :

Unity brain name: BananaBrain
        Number of Visual Observations (per agent): 0
        Vector Observation space type: continuous
        Vector Observation space size (per agent): 37
        Number of stacked Vector Observation: 1
        Vector Action space type: discrete
        Vector Action space size (per agent): 4
        Vector Action descriptions: , , , 
Timing Atk Ratio: 10%
Using CEVAE_QNetwork.
Unity Worker id: 10  T: 1  Use baseline:  0  CEVAE:  1
With:  9.67 % timing attack
Episode 0   Score: 0.00, Average Score: 0.00
With:  11.0 % timing attack
Episode 5   Score: 1.00, Average Score: 0.17
With:  11.33 % timing attack
Episode 10   Score: 0.00, Average Score: 0.36
With:  10.33 % timing attack
Episode 15   Score: 0.00, Average Score: 0.56
...
Episode 205   Score: 10.00, Average Score: 9.25
With:  9.33 % timing attack
Episode 210   Score: 9.00, Average Score: 9.70
With:  9.0 % timing attack
Episode 215   Score: 10.00, Average Score: 11.10
With:  8.33 % timing attack
Episode 220   Score: 14.00, Average Score: 10.85
With:  12.33 % timing attack
Episode 225   Score: 19.00, Average Score: 11.70
With:  11.0 % timing attack
Episode 230   Score: 18.00, Average Score: 12.10
With:  7.67 % timing attack
Episode 235   Score: 21.00, Average Score: 11.60
With:  9.67 % timing attack
Episode 240   Score: 16.00, Average Score: 12.05

Environment solved in 242 episodes!     Average Score: 12.50
Environment solved in 242 episodes!     Average Score: 12.50 +- 4.87
############# Basic Evaluate #############
Using CEVAE_QNetwork.
Evaluate Score : 12.6
############# Noise Evaluate #############
Using CEVAE_QNetwork.
Robust Score : 12.5

Reference

This fun work was initialzed when Danny and I first read the Causal Variational Model between 2018 to 2019 with the helps from Dr. Yi Ouyang and Dr. Pin-Yu Chen.

Please consider to reference the paper if you find this work helpful or relative to your research.

A non-archival and preliminary venue was presented in ICLR 2021 Self-supervision for Reinforcement Learning Workshop, spotlight, long contributed talk. We appreciate the research community and the orgainzers.

@article{yang2021causal,
  title={Causal Inference Q-Network: Toward Resilient Reinforcement Learning},
  author={Yang, Chao-Han Huck and Hung, I and Danny, Te and Ouyang, Yi and Chen, Pin-Yu},
  journal={arXiv preprint arXiv:2102.09677},
  year={2021}
}

Training a Resilient Q-Network against Observational Interference, Causal Inference Q-Networks

Related tags

Overview

Obs-Causal-Q-Network

Environment Setup

1. Example of Training Causal Inference Q-Network (CIQ) on Cartpole

2. Example of Training a "Variational" Causal Inference Q-Network on Unity 3D Banana Navigator

Reference

Owner

Official implementation of VQ-Diffusion

A 3D sparse LBM solver implemented using Taichi

Flexible time series feature extraction & processing

Trading and Backtesting environment for training reinforcement learning agent or simple rule base algo.

Voice Conversion by CycleGAN (语音克隆/语音转换)：CycleGAN-VC3

Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting

Put blind watermark into a text with python

Weighted K Nearest Neighbors (kNN) algorithm implemented on python from scratch.

Code for our SIGCOMM'21 paper "Network Planning with Deep Reinforcement Learning".

Domain Adaptation with Invariant RepresentationLearning: What Transformations to Learn?

[NeurIPS 2021] “Improving Contrastive Learning on Imbalanced Data via Open-World Sampling”,

Gauge equivariant mesh cnn

[WWW 2022] Zero-Shot Stance Detection via Contrastive Learning

Code for Transformers Solve Limited Receptive Field for Monocular Depth Prediction

ESTDepth: Multi-view Depth Estimation using Epipolar Spatio-Temporal Networks (CVPR 2021)

Tools for manipulating UVs in the Blender viewport.

wgan, wgan2(improved, gp), infogan, and dcgan implementation in lasagne, keras, pytorch

Code release for "Masked-attention Mask Transformer for Universal Image Segmentation"

Extracting knowledge graphs from language models as a diagnostic benchmark of model performance.

Implementation of gMLP, an all-MLP replacement for Transformers, in Pytorch