Skip to content

sannebh/riskaverse_rl

Repository files navigation

Risk averse reinforcement learning.

This code combines recent progress in Distributional Reinforcement Learning with Quantile Regression (https://arxiv.org/pdf/1710.10044.pdf) with classic theory to create risk averse, safe algorithm.

To download the code:

git clone https://github.com/sannebh/riskaverse_rl/

The cartpole environment is implemented using OpenAI Gym. This is a toolkit for developing and comparing reinforcement learning algorithms. Installation guidelines and other information can be found at https://github.com/openai/gym.

We use Pytorch to build the networks, for installation and other information please see https://pytorch.org/.

The common folder contains helper functions and classes: layers, replay buffert and wrappers. They are all from Open AI Baselines (https://github.com/openai/baselines).

The risk averse strategies can be found in risk_strategies.py. wind_world.py and contains the Windy World environment.

qr-dqn_cart and qr-dqn_windy contains the implementations themselves, in Cartpole Environment and Windy Gridworld environment.

About

Risk averse distributional reinforcement learning applied in cart pole and windy world environment

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages