Dynamic Bottleneck

Introduction

This is a TensorFlow based implementation for our paper on

"Dynamic Bottleneck for Robust Self-Supervised Exploration". NeurIPS 2021

python3.6 or 3.7, tensorflow-gpu 1.x, tensorflow-probability, openAI baselines, openAI Gym

The following command should train a pure exploration agent on "Breakout" with default experiment parameters.

python run.py --env BreakoutNoFrameskip-v4

The following command should train a pure exploration agent on "Breakout" with randomBox noise.

python run.py --env BreakoutNoFrameskip-v4 --randomBoxNoise

The following command should train a pure exploration agent on "Breakout" with Gaussian noise.

python run.py --env BreakoutNoFrameskip-v4 --pixelNoise

The following command should train a pure exploration agent on "sticky Breakout" with a probability of 0.25

python run.py --env BreakoutNoFrameskip-v4 --stickyAtari

ICM: We use the official code of "Curiosity-driven Exploration by Self-supervised Prediction, ICML 2017" and "Large-Scale Study of Curiosity-Driven Learning, ICLR 2019".
Disagreement: We use the official code of "Self-Supervised Exploration via Disagreement, ICML 2019".
CB: We use the official code of "Curiosity-Bottleneck: Exploration by Distilling Task-Specific Novelty, ICML 2019".

Name		Name	Last commit message	Last commit date
Latest commit History 1 Commit
README.md		README.md
__init__.py		__init__.py
cnn_policy.py		cnn_policy.py
cppo_agent.py		cppo_agent.py
dynamic_bottleneck.py		dynamic_bottleneck.py
mpi_utils.py		mpi_utils.py
para.json		para.json
recorder.py		recorder.py
rollouts.py		rollouts.py
run.py		run.py
utils.py		utils.py
vec_env.py		vec_env.py
wrappers.py		wrappers.py