Authors: Ibrahim El Shar and Daniel Jiang

This is the source code for our paper "Lookahead-bounded Q-learning" published at ICML 2020.

We propose a new provably convergent variant of Q-learning that leverages upper and lower bounds derived using information relaxation techniques to improve the performance of standard Q-learning.

You can use the following Bibtex entry to cite our paper:

  @article{elshar2020lookahead,
    Title = {Lookahead-bounded Q-learning},
    Author = {Ibrahim El Shar and Daniel Jiang},
    journal={Proceedings of the 37th International Conference on Machine Learning (ICML)},
    Year = {2020},
    address={Vienna, Austria}
  }

Code was tested on Python version 3.6

Build a working python enviromnent for e.g. using Anaconda.

Install packages from requirements.txt

Clone this repo:

	git clone https://github.com/ibrahim-elshar/LBQL_ICML2020.git

There are five environments organized in folders inside src: Windy Gridworld (WG), Stormy Gridworld (SG), Repositioning for Car-sharing in 2 stations Platform (2-CS-R), Pricing for Car-sharing in 2 stations Platform (2-CS) and Pricing for Car-sharing in 4 stations Platform (4-CS). Each environment is placed in a separate folder.

Each folder contains:

• Environment file, e.g. carsharing.py for 2-CS. Running this file will produce the optimal Q-value Qstar.pkl if applicable.
• agents.py file that contains the code for QL, Double-QL, SQL, BCQL and LBQL algorithms.
• run.py file which re-runs the experiments for the environment and reproduce the performance and relative error plots.

Hyperparameters for an algorithm can be set by chaninging the corresponding class default parameters in agents.py file.

python agents.py will produce LBQL vs QL bounds plots.

To rerun the experiments for an evironment cd to the environment folder first then:

$ python run.py