CS7642-Reinforcement Learning and Decision Making Solved

1.1     Description
For this project, you will be implementing and training an agent to successfully land the “Lunar Lander” (LunarLander-v2) in OpenAI gym. You are free to use and extend any type of RL agent.

1.2      Lunar Lander Environment
In this problem you will be working in an 8-dimensional state space, with six continuous state variables and two discrete ones. The action space is discrete. The four discrete actions available are: do nothing, fire the left orientation engine, fire the main engine, fire the right orientation engine. The landing pad is always at coordinates (0,0). Coordinates consist of the first two numbers in the state vector. The base reward for moving from the top of the screen to the landing pad depends on a number of factors. If the lander moves away from the landing pad it is penalized the amount of reward that would be gained by moving towards the pad. An episode finishes if the lander crashes or comes to rest, receiving an additional -100 or +100 points, respectively, on top of the base reward. Each leg-ground contact is worth +10 points. Firing the main engine incurs a -0.3 point penalty and firing the orientation engines incur a -0.03 point penalty, for each occurrence. Landing outside of the landing pad is possible. Fuel is infinite, so, an agent could learn to fly and land on its first attempt. The problem is considered solved when achieving a score of 200 points or higher on average over 100 consecutive runs. You are encouraged to look at the Lunar Lander source code for more information on the environment and reward structure.

1.3      State Representation
As noted earlier, there are four actions available to your agent: do nothing, fire the left orientation engine, fire the main engine, fire the right orientation engine. Additionally, at each time step, the state is provided to the agent as an 8-tuple:

                                                                                    (x,y,x,˙ y,θ,˙  θ,leg˙  L,legR)

The state variables x and y are the horizontal and vertical position, ˙x and ˙y are the horizontal and vertical speed, and θ and θ˙ are the angle and angular speed of the lander. Finally, legL and legR are binary values to indicate whether the left leg or right leg of the lunar lander is touching the ground.

1.4     Procedure
This problem is more sophisticated than anything you have seen so far in this class. Make sure you reserve enough time to consider what an appropriate solution might involve and, of course, enough time to build it.

•    Create an agent capable of solving the Lunar Lander problem found in OpenAI gym

–    Upload/maintain your code in your private repo at https://github.gatech.edu/gt-omscs-rldm

–    Use any RL agent discussed in the class as your inspiration and basis for your program

•    Create graphs demonstrating

–    The reward for each training episode while training your agent

–    The reward per episode for 100 consecutive episodes using your trained agent

–    The effect of at least 3 hyper-parameters of your choosing on your agent ∗ You will select the ranges to be evaluated.

1

∗ Evaluate your results in the context of the algorithm and the problem. Why do your results make sense?

–    Anything else you may think appropriate

•    We’ve created a private Georgia Tech GitHub repository for your code. Push your code to the personal repository found here: https://github.gatech.edu/gt-omscs-rldm

•    The quality of the code is not graded. You don’t have to spend countless hours adding comments, etc. But, it will be examined by the TAs.

•    You may use any function approximator to solve this problem, however we strongly recommend you use a non-linear function approximator such as a neural network. If you decide to train a neural network, you must train it using PyTorch as an automatic differentiation tool. Other neural network training libraries such as Tensorflow, Keras, Theano are not allowed. PyTorch is well liked amongst researchers for its pythonic feel, and we will hold a separate office hours to do a quick overview of PyTorch.

•    Make sure to include a README.md file for your repository

–    Include thorough and detailed instructions on how to run your source code in the README.md

–    If you work in a notebook, like Jupyter, include an export of your code in a .py file along with your notebook

–    The README.md file should be placed in the project 2 folder in your repository.

•    You will be penalized by 25 points if you:

–    Do not have any code or do not submit your full code to the GitHub repository

–    Do not include the git hash for your last commit in your paper

•    Write a paper describing your agent and the experiments you ran

–    Include the hash for your last commit to the GitHub repository in the header on the first page of your paper.

–    Make sure your graphs are legible and you cite sources properly. While it is not required, we recommend you use a conference paper format. For example https://www.ieee.org/conferences/ publishing/templates.html

–    5 pages maximum – really, you will lose points for longer papers.

–    Explain your experiments.

–    Graph: Reward at each training episode while training your agent and discussion of results

–    Graph: Reward per episode for 100 consecutive episodes using you trained agent and discussion of the results.

–    Graph: Effect of hyperparameters and discussion of the results.

–    Explanation of algorithms used: what worked best? what didn’t work? what could have worked better?

–    Justify your choices

∗ Unlike project 1 there are multiple ways of solving this problem and you have a lot of discretion over the general approach you take as well as specific design decisions. Explain to the reader why, from amongst the multiple alternatives, you chose the ones you did.

∗ Same applies to hyperparameters experimentation – Explanation of pitfalls and problems you encountered.

–    What would you try if you had more time?

–    Save this paper in PDF format.

–    Submit to Canvas!

•    Using a Deep RL library instead of providing your own work will earn you a 0 grade on the project and you will be reported for violating the Honor Code.

1.5     Notes
•    If you get a Box2D error when running gym.make(‘LunarLander-v2’), you will have to compile Box2D from source. Please follow these steps and try running Lunar Lander again https://github.com/openai/ gym/issues/100:

pip uninstall box2d-py git clone https://github.com/pybox2d/pybox2d cd pybox2d/ python setup.py clean python setup.py build sudo python setup.py install

•    Windows users may install Box2D with the prebuilt Python wheels provided at https://www.lfd.uci. edu/~gohlke/pythonlibs/#pybox2d

•    Even if you don’t build an agent that solves the problem you can still write a solid paper.

1.6     Resources
1.6.1       Lectures

•    Lesson 8: Generalization

1.6.2       Readings

•    Chapter 9 (On-Policy Prediction with Approximation) of Sutton and Barto 2020

1.7      Source Code
•    https://github.com/openai/gym/blob/master/gym/envs/box2d/lunar_lander.py

1.7.1        Documentation

•    http://gym.openai.com/docs/