Deep Q-Learning Networks for Atari Breakout game
Budget: $10 – $63 USD
This Task is to implement Deep Q-Learning Networks (DQN) for an RL agent that lean to play Atari
games, in particular the game of Breakout.
Your Task:
Your task is to implement (at least) ONE DQN agent, report the performances with visualizations, and discuss
the results in depth.
Requirements:
The agent should utilize a Deep Q-learning network as the model.
The agent must use TWO networks -- a prediction and a target networks.
As a hint, you can call get_weights() and set_weights() in the networks to transfer weights.
Also it should use a reply/experience memory.
Train the agent/model using various epsilon values.
Try different model/training hyperparameters and tune them, as usually done in Deep Learning.
A. DQN Agents
The agent must be a Convolutional Neural Network (CNN). It receives an observation, which represents
a color image of 210 x 160 x 3 pixels in a numpy array, and produces an output of length
env.action_space.n (which is 4 for the Breakout's reduced action space) that indicates the probability
distribution of the actions. All other details of the model are up to you.
You first run the model for a sufficient number of times to populate the reply memory. Then you train
the model, using the epsilon too, and tune the model and training hyperparameters such as learning
rate.
Here are some ideas for challenge if you are interested. You figure out how to implement
by yourself.
Stack 4 frames and use it as the input to the QNetwork.
Use a feed-forward network (and the gray scale image).
B. Performance Report
Show charts that display the effect of epsilon as well.
Show a video of the best run.
2. Write-up report.
Minimum 2.5 pages (in pdf of docx).
Description of your deep network models and hyperparameter settings.
Your comments on DRL agents after learning. How long did it take to train? Did the learned agent
play 'well'? Did it exceed your expectation? Write some observations of the agent learning and/or
behavior.
Write your general reflections on the problem:
What you learned from this exercise.
How difficult you felt this exercise was.
Any particular difficulties you encountered.
How you would do/approach differently next time (if there was one).
and anything else.
games, in particular the game of Breakout.
Your Task:
Your task is to implement (at least) ONE DQN agent, report the performances with visualizations, and discuss
the results in depth.
Requirements:
The agent should utilize a Deep Q-learning network as the model.
The agent must use TWO networks -- a prediction and a target networks.
As a hint, you can call get_weights() and set_weights() in the networks to transfer weights.
Also it should use a reply/experience memory.
Train the agent/model using various epsilon values.
Try different model/training hyperparameters and tune them, as usually done in Deep Learning.
A. DQN Agents
The agent must be a Convolutional Neural Network (CNN). It receives an observation, which represents
a color image of 210 x 160 x 3 pixels in a numpy array, and produces an output of length
env.action_space.n (which is 4 for the Breakout's reduced action space) that indicates the probability
distribution of the actions. All other details of the model are up to you.
You first run the model for a sufficient number of times to populate the reply memory. Then you train
the model, using the epsilon too, and tune the model and training hyperparameters such as learning
rate.
Here are some ideas for challenge if you are interested. You figure out how to implement
by yourself.
Stack 4 frames and use it as the input to the QNetwork.
Use a feed-forward network (and the gray scale image).
B. Performance Report
Show charts that display the effect of epsilon as well.
Show a video of the best run.
2. Write-up report.
Minimum 2.5 pages (in pdf of docx).
Description of your deep network models and hyperparameter settings.
Your comments on DRL agents after learning. How long did it take to train? Did the learned agent
play 'well'? Did it exceed your expectation? Write some observations of the agent learning and/or
behavior.
Write your general reflections on the problem:
What you learned from this exercise.
How difficult you felt this exercise was.
Any particular difficulties you encountered.
How you would do/approach differently next time (if there was one).
and anything else.