Skip to content

In this project, I use transfer learning on a cartpople (openAI) env to demonstrate the transferring of weights in Reinforcement learning and its advantages and limitations.

License

Notifications You must be signed in to change notification settings

anksng/Transfer-learning-in-Deep-Reinforcement-learning

Repository files navigation

Transfer-learning-in-Deep-Reinforcement-learning

In this project, I use transfer learning on cartpole (openAI) env to demonstrate the transferring of weights in Reinforcement learning and its advantages and limitations.

Outline of this project:

  1. We train a basic cartpole using the data generated by playing "some random games"
  2. Later transfer the knowledge(weights) learnt from this to another similar task or environment.
  3. New task(env) is created by changing the source file of cart pole(details below)

Version details :

  • Cartpole-normal.ipnyb - This is the basic cartpole with default parameters
  • Cartpole-v1.ipnyb - This cartpole is of different dimensions and gravity and trained without using transfer learning
  • Cartpole-v2.ipnyb - This is same as v2 but training is done with transfering previous knowledge learned from normal Cartpole

To change the dimensions and env, open the source file from openai alt text

Results :

My observations

alt text

Cartpole source code which i used had the following values:

Although it can be freely changed as this is an experiment to create a similar env with different dynamics (E.g. in second cofig , gravity is 24 which is same as Jupiter. Thus imagine training a cart to balance a pole on Earth and then use it to work on Jupiter!)

alt text

alt text

Final results :

alt text

alt text

Watch the difference in performance here -

v1-

https://youtu.be/ImTZXh1SDQ8

v2- with Transferring weights

https://youtu.be/h_TIQVivaeM

About

In this project, I use transfer learning on a cartpople (openAI) env to demonstrate the transferring of weights in Reinforcement learning and its advantages and limitations.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published