Beyond Prioritized Experience Replay

Ablation study of the Prioritized Experience Replay presentend by Schau et al.

Prerequisite

• Pytorch
• Numpy
• random
• Gym
• Matplotlib
• pyvirtualdisplay
• JSAnimation.IPython_display
• IPython
• python-opengl
• ffmpeg
• xvfb

Run

To replicate the results, we suggest running the code on COLAB in the same way as it is run in Tests.ipynb

The code in this repo is based on the following repo:

• Rainbow is all you need

Here are some of the results we got:

• Performance for priority-based PER in different environments:

• Performance for priority-based PER with different memory size(Cartpole v0)

• Performance for priority-based PER with different stalness Coeff (Cartpole v0)

• Performance for rank-based PER with different stalness Coeff (Cartpole v0)

• Performance for priority-based PER with different Positive penalty coefficients (Cartpole v0)

• Performance for rank-based PER with different positive penalty coeff (Cartpole v0)

• Performance for priority-based PER without and with differential method (Cartpole v0)

• Performance for rank-based PER without and with differential method (Cartpole v0)

• Performance for the 3 different types of PER (priority-based, rank-based and hybrid approach comibining positive penalty and staleness)(Cartpole v0)