No description, website, or topics provided.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
Final_Report.pdf
PyTorch_DQN_SuperMarioBros_8screen_CNN.py
PyTorch_DQN_SuperMarioBros_8screen_NN.py
README.md
test.py

README.md

Training-DQN-to-play-Super-Mario-Bros.

We present a deep learning model to successfully learn control policies from high-dimensional input data using reinforcement learning. The model is based on the idea of Deep Q-Network (DQN), with convolutional neural network trained by Q-learning algorithm, whose input is tile representation of the screen and output is a value estimation function. Also, replay buffer, target network and double Q-learning are applied to lower data dependency and approximate real gradiant descent. We applied our model to Super Mario Bros., and get some preliminary results.

To run this code, you need to add the following environment into OpenAI Gym: https://github.com/ppaquette/gym-super-mario

Video: https://youtu.be/CX4pdlcjxPQ

Author = {Guanya Shi, Botao Hu, Yan Wu}

Email = {gshi, bhu2, ywu5}@caltech.edu