Druce Rl Save

Notebooks and code for Alpha Architect post on reinforcement learning.

• Tic-Tac-Toe.ipynb - Table-based reinforcement learning to play Tic-Tac-Toe, and analogous if pointless deep learning algo
• Cart-Pole.ipynb - Building deep reinforcement learning algos from scratch with Keras for OpenAI environments like Cartpole and LunarLander.
  • DQN
  • Policy Gradient (REINFORCE)
  • REINFORCE with baseline
  • Run_CartPole.ipynb, Run_LunarLander.ipynb - only run saved good models, don't train
• Ray_tune.ipynb - Similar but with state of the art RL from UC Berkeley Ray project
• Trading_with_RL.ipynb - Algos to trade fake market data, inspired by Gordon Ritter paper Machine Learning for Trading. This should run in Google Colab.

Typical installation procedure:

• Install Anaconda python data science distribution

• Make an environment like

  conda create --name tf tensorflow
  

  or if you have Nvidia GPU

  conda create --name tf_gpu tensorflow-gpu 
  

  This should install requirements like working Nvidia drivers

• Upgrade TensorFlow to latest version with

  pip install --upgrade tensorflow
  

• Install additional requirements as necessary - requirements.txt has python modules installed at time of testing.

  pip install -r requirements.txt
  

• TensorFlow Docker install may also be a good way to start but has not been tested.

• Run notebooks using

  jupyter notebook