Reinforcement Learning - Assignment

Work in progress - please check back for updates.

Part 1 - Improve Cart-Pole

If you've not already, complete through goal 2 of Coding Challenge 1. The first part of the assignment is to improve the policy for the agent that is trying to keep the pole upright on the cart in the notebook 16_reinforcement_learning.ipynb. Read the notebook through A simple hard-coded policy, and then add your own code (in your local Docker-powered environment) to try to do better.

As noted, the main problem with the simple approach is that it ends up moving the cart too abruptly and losing stability. You can refer to the later sections of the notebook for some hints on ways to improve with various techniques, but your goal here is to write your own code leveraging knowledge from the various resources you have available.

Please turn in an updated notebook with your strategy, as well as text in the notebook describing it and its performance.

Part 2 - Play Pong

Implement an agent to play Pong - the OpenAI Gym integrates an Atari environment, and you can start by trying example agents. The random agent is a simple initial example, and the keyboard agent can allow you to actually play the game yourself (and potentially use that data to train some other agent). The example notebook from the first coding challenge has useful snippets of code for visualizing what is going on.

After getting set up and playing with a few agents, check out this article on using policy gradients to train a model to play Pong. Code your own agent based on this technique - this video illustrates the end result of such an agent.

Part 3 - An environment of your choice

OpenAI supports many interesting environments - pick one, and start by trying to run existing agents and find prior work (as with Pong). Then take a shot at writing your own - this is open-ended, so have fun and explore something that interests you, but also chat and share what you're doing so people can collaborate and learn from one another.