yixqiao/GameAI

Neural network trained evolutionarily to play a game

GameAI

Neural network trained evolutionarily to play a game.
Made with Java and Processing.

Demo

This is a demo of the AI playing one game.

Quickstart

1. Clone repository
2. Run ./gradlew build && ./gradlew run
3. You will see the AI play a random game. Press space to restart a new game once a game ends.

Playing & Training

1. Open src/main/java/gameai
2. In the first line of main, there should be a string option
3. Set that to play to play the game as a person, or to train to train
   • To play, use left, right, and down arrow keys
   • In training, press space to view the next generation play, and then space again to go back to silent training

Game & AI info

Game:

• Move left, right
• Stomp opponenets
• Death if touches opponents
• Score increases every second and stomping opponents
  • Stomping increases by a "streak" mechanic: more stomps without touching ground = higher score
  • Stomping without hitting opponents loses a few points

AI:

• Recurrent Neural Network
  • Fully implemented in pure Java all the way down because I couldn't find any well-documented matrix libraries
  • Two hidden layers (sizes of 64 & 16)
  • Recurrent layer of size 10
  • Input:
    • Closest 6 enemies: x, y, m speed
    • Player: x, y, m speed, isStomping
  • Output:
    • Left, right, down