- A Deep Convolutional Network built with Tensorflow is used to get environment state from raw image data from Chrome Browser, processed using OpenCV.
- An OpenAI gym environment had been modified to be able to connect with browser game and training notebook. From the predicted state variables, the agent explores the environment with epsilon-greedy policy.
- To increase sample efficiency and independently distribute the states for training, Experience Replay had also been used.
- Start by cloning the repository.
$ git clone https://github.com/ShivenTripathi/Hack2Learn.git
- Install gym-env by following instructions in Hack2Learn/gym-chrome-dino/
- Use Hack2Learn/DQN Agent to play Chrome Dino.ipynb to train, test your model.
- You can save and load the model to view result from the the last cells.
- Use the requirements.txt file
conda install --file requirements.txt
Hack-2-Learn will hold short and interesting projects that can be completed in 1-2 weeks.
