Reinforcement Learning Summer School 2017

To compile the tile-coding library:

1. cd tile-coding
2. make
3. export PYTHONPATH=$PYTHONPATH:pathto/tile-coding

300+ Notebooks from McGill COMP-767, Intro to RL

A collections of notebooks written by the students of McGill COMP-767, Intro to RL. We had a "bring your own assignment" model in which the students would create their own "assignment" related to the course material. The assignments would generally take the form of a Jupyter notebook exploring some questions empirically and/or theoretically.

• Course schedule and resources
• Github Classroom containing the notebooks

With no particular order, a few awesome notebooks :

• Empirical study of $ETD(\lambda,\beta)$
• An Empirical Exploration of Provisional Temporal Difference Learning
• Synthetic gradients for REINFORCE
• A tile coder in theano for Reinforcement Learning tasks
• An implementation of LSTD(lambda), TD(0) and RLSTD(0) for the Boyan Chain problem
• Bias-Variance trade-off for Monte-Carlo methods
• An analysis of bias-variance tradeoff of Sarsa, Expected Sarsa, Double Sarsa, and Double Expected Sarsa
• Convergence visualization in the eigenbasis of $P_\pi$

Wrapper to Marlos Machado's Linear Features for ALE

The instructions are provided in the README.md in :

• shallowpy

The example code relies on memory overcommitment which is rather useful to know about. The overcommit mode can be read/set via cat /proc/sys/vm/overcommit_memory. From the Kernel documentation :

1 - Always overcommit. Appropriate for some scientific applications. Classic example is code using sparse arrays and just relying on the virtual memory consisting almost entirely of zero pages.

Other References

• OpenAI Baselines
• The original Mountain Car code written by Richard Sutton Mountain Car Software
• The tile coding library used by the RLAI Tile Coding Software