/tlrl

Primary LanguageJupyter Notebook

Introduction

  • This repository is our work for our group project for the grad course CSC2541 taken at the University of Toronto.
  • We attempt to improve existing Intertask Mapping methods for transfer in reinforcement learning, specifically Taylor's MASTER and Ammar's TrRBM.

Group Members

  • Daniel Goldberg, Jeremy Ma, Irene Jiang, Feng Chi and Akshay Budhkar.

Ammar TrRBM extension

  • We use TrRBM in a model-free setting.

Taylor MASTER extension

  • We use MASTER in an autonomous setting.

Experiments

  • We run our experiments on the canonical 2D -> 3D Mountain car problem as well as four other transfer scenarios.
  • We designed new environments (3D Mountain car and 3D Cart Pole) to work in lieu with the existing Open AI environments.

Acknowledgements

The group would like to acknowledge the direction it received from Professor Grosse during the course of the term. We would also like to thank Matthew Taylor, Haitham Bou Ammar, Denny Britz, Pieter Abbeel and Andrew Ng for indirectly motivating us to pursue this project. Finally, we would like to acknowledge Open AI for providing an intuitive and fairly easy-to-use library for running our experiments.