Seasons Of Code

Learn To Flap    • Rohit Kumar Jena    • Chitwan Saharia   

Learn To Flap

Learn To Flap

Learn to Flap is a machine learning project to use supervised and unsupervised learning algorithms to build an AI that plays the game “Flappy Bird”. We plan to start with using supervised learning algorithms by using Support Vector Classifier, Neural Networks, and also use unsupervised methods like Q-learning.

We may also work on other variations of the game if time permits.
Reading material would include some book on machine learning to get the basics and some work that has been done by other people (will provide them as the project progresses). Refer to this link for a good deep learning tutorial.

Week 1

  • Learnt basic methods of machine learning like Linear and Logistic Regression
  • Enrolled for the ML course by Andrew Ng on Coursera
  • Read up multiclass logistic regressions and did a few tutorials

Week 2

  • Implemented a tensorflow tutorial
  • Started reading up Neural Networks theory from various sources like Andrew Ng’s Course

Week 3

  • Started implementing stuff learnt in the past couple of weeks on Flappy Bird clone
  • Decided attributes for feature space
  • Finished upto Week 5 in Andrew Ng’s course

Week 4

  • Used the above model to train neural Network
  • Started collecting data for Flappy Bird using a game playing bot

Week 5 - 8

  • Spent on implementing the neural network, debugging and improving
    • Implemented the neural network using keras and got a good accuracy
    • Increased neurons in the first hidden layer to improve performance
    • made a few calibrations to remove abrupt gap in the beginning
    • Increased the depth of the neural net to three layers
    • Increased size of training set

A student is expected to have interest in machine learning and AI (coding experience isn’t necessary).