Flappy Bird RL by djbyrne - 1

MobileGames & ProjectsFrameworks

A reinforcement learning environment based on the mobile game "Flappy Bird" built using the Unity ml-agents framework

Unknown VersionUnknown LicenseUpdated 1 day agoCreated on May 14th, 2019
Go to source

FlappyBirdRL

Reinforcement Learning environment for the game Flappy Bird using the Unity ML Agents library

Requirements

  • Unity version 2019.1.0f2
  • ml-agents 0.8.1

Action Space

  • 1 discrete action “flap”: 1 = move up, 0 = do nothing (fall down) Note: once the agent has “flapped” they must do nothing for the next action. This simulates the clicking functionality of the original game

State Space

State space comprises of 6 feature vectors normalized between [0,1]

  • agent height
  • agent Y velocity
  • last action taken
  • height of the next top pipe
  • height of the next bottom pipe
  • X distance to next pipe

Rewards

The goal is to avoid the pipes for as long as possible

  • +0.1 for each timestep where the agent is alive
  • +1.0 for each pipe passed successfully
  • -1.0 for each collision and terminated

Info

The game runs at x20 normal speed in order to speed up training.

Show all projects by djbyrne