#組員:張嘉哲、林暘竣

Please complete each homework for each team, and
mention who contributed which parts in your report.

In this assignment, we will solve the classic control problem - CartPole.

CartPole is an environment which contains a pendulum attached by an un-actuated joint to a cart, and the goal is to prevent it from falling over. You can apply a force of +1 or -1 to the cart. A reward of +1 is provided for every timestep that the pendulum remains upright.

• OpenAI gym
• TensorFlow
• Numpy
• Scipy
• IPython Notebook

If you already have some of above libraries installed, try to manage the dependencies by yourself.

If you are using a new environment (may be virtual), the preferred approach for installing above dependencies is to use Anaconda, which is a Python distribution that includes many of the most popular Python packages for science, math, engineering and data analysis.

1. Install Anaconda: Follow the instructions on the Anaconda download site.
2. Install TensorFlow: See anaconda section of TensorFlow installation page.
3. Install OpenAI gym: Follow the official installation documents here.

If you are unfamiliar with Numpy or IPython, you should read materials from CS231n:

• Numpy tutorial
• IPython tutorial

Also, knowing the basics of TensorFlow is required to complete this assignment.

For introductory material on TensorFlow, see

• MNIST For ML Beginners from official site
• Tutorial Video from Stanford CS224D

Feel free to skip these materials if you are already familiar with these libraries.

1. Start IPython: After you clone this repository and install all the dependencies, you should start the IPython notebook server from the home directory
2. Open the assignment: Open HW2_Policy_Graident.ipynb, and it will walk you through completing the assignment.

• [+20] Construct a 2-layer neural network to represent policy

• [+30] Compute the surrogate loss

• [+20] Compute the accumulated discounted rewards at each timestep

• [+10] Use baseline to reduce the variance

• [+10] Modify the code and write a report to compare the variance and performance before and after adding baseline (with figures is better)

• [+10] In function process_paths of class PolicyOptimizer, why we need to normalize the advantages? i.e., what's the usage of this line:

  p["advantages"] = (a - a.mean()) / (a.std() + 1e-8)

  Include the answer in your report

• Office hour 2-3 pm in 資電館 with YenChen Lin.
• Due on Oct. 17 before class.