Machine Learning

Syllabus

Git & GitHub: proficiency assumed
- Otherwise: Git & GitHub
Jupyter Notebook [1w]
- crash course: Edit, run, publish
- Login to shared notebook
- setup your own server
Q-Learning Intro [3w]
- Along the lines of Microsoft ML Chapter 8
- Assumes we can observe the entire state.
- Reward / action only depends on the current state of the game.
- Immediate reward given by game score (points)
- Q-Function is a table (Q-Table) dynamically updated by Bellman-formula
Hidden Markov Models (a tiny bit of theory)
- → what if we cannot observe all the internal state?
- → infer probability of hidden (latent) state from observations
Deep Reinforcement Learning
- Q-Function is not a table, but a neural network, which is trained by repeatedly running the game and updating the neuron states.
- builds on top of sca's NN program
- Prereqs:
  - OpenAI Gym environment: https://gym.openai.com/docs/
  - keras / tensorflow?