I am implementing a lot of reinforcement learning and imitation learning algorithms since I'm sick of reading about them but not really understanding them.
Python Shell
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
dqn Adding Enduro results. Stuff I need to do: Apr 9, 2017
q_learning Update README.md Jan 3, 2017


Reinforcement Learning Algorithms

I will use this repository to implement various reinforcement learning algorithms (also imitation learning), because I'm sick of reading about them but not really understanding them. Hence, hopefully this repository will help me understand them better. I will also implement various supporting code as needed, such as for simple custom scenarios like GridWorld. Or I can use OpenAI gym. Click on the links to get to the appropriate algorithms. Each sub-directory will have its own READMEs with results there, along with usage instructions.

Here are the algorithms currently implemented or in progress:

Note: "Vanilla Policy Gradients" refers to the REINFORCE algorithm, also known as Monte Carlo Policy Gradient. Sometimes it's called an actor-critic method and other times it's not. Even if it's considered an actor-critic method, the usual way we think of actor-critic involves a TD update rather than waiting until the end of an episode to get returns.


Right now the code is designed for Python 2.7, but it should be compatible with Python 3.5+, with the possible exception of if the bash scripts can't tell the difference between which Python versions I'm using.

In short:

  • Python 2.7.x
  • Tensorflow 1.2.0

GPU and TensorFlow

(Update 06/16/17, these are out of date ... just install with pip and preferably virtualenv. It's so much easier.)

I installed TensorFlow 1.0.1 from source. For the configuration script, I used CUDA 8.0, cuDNN 5.1.5, and compute capability 6.1.

Compiling from source means I can get faster CPU instructions. This requires bazel plus extra compiler options. I used:

bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

This resulted in ton of warning messages but I ended up with:

Target //tensorflow/tools/pip_package:build_pip_package up-to-date:
INFO: Elapsed time: 884.276s, Critical Path: 672.19s

and things seem to be working. Then run the command:

bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

To get a wheel, which we then do a pip install. But be careful due to pip on anaconda vs pip with default python. I use anaconda. And make sure you're not in either the tensorflow or the bazel directories!

Track the GPU usage with nvidia-smi. Unfortunately, that's only for one time-step, but we can instead run:

while true; do nvidia-smi --query-gpu=utilization.gpu --format=csv >> gpu_utilization.log; sleep 10; done;

Or something like that. It will record the output every 10 seconds and dump it into the log file. Ideally, GPU usage should be as high as possible (100% or close to it).


I have read a number of reinforcement learning paper references to help me out. A list of papers and summaries (for a few of them) are in my paper notes repository.