jamro / neuraltank

Neural Network for JsBattle

NeuralTank

The project is about evolving a neural network that successfully controls a tank and fights in JsBattle coding game.

Neural Network

Inputs

• wallDistance - distance to the nearest wall in radar range. The value is in range of -1 to 1. The input is set to 1 if there is no wall in radar range.
• enemyDistance - distance to the nearest enemy in radar range. The value is in range of -1 to 1. The input is set to 1 if there is no enemy in radar range.
• enemyAngle - how far away the enemy is from the center of radar beam. -1 represents the left edge, 0 is the center and -1 is the right edge.
• collision - determine whether the tank hit an obstacle recently. 1 means recent collision and 0 is opposite
• boost - the amount of boost left (a number between -1 and 1)
• targetting alarm - determine whether the tank is in radar beam of an enemy (0 or 1)

Output

• throttle - throttle controls. Negative numbers result in backward movement while positive values move the tank forward
• turn - the speed of turning
• shoot - determine whether to shot or not.
• shoot - determine whether to boost or not.

Network Structure:

• Six inputs
• Two hidden layers (7 and 6 nodes)
• Four outputs

Evolution

The neural network is evolved by a genetic algorithm. The Fitness formula is based on the score achieved during the battle. Rendering could be disabled to speed up evolution. Each step of the simulation is saved in Local Storage so simulation could be stopped and continued later

Installation

1. Download sources from GitHub
2. Install all dependencies by executing npm i in the project folder
3. Run npm start and visit http://localhost:9000

Import population

Evolving neural network takes time. There is a snapshot of 1000th generation for the impatient. Download the neuraltank-gen1000-size200.json file and import it to skip 1K of generation steps.