the-binding-of-isaac-in-js-ai-bot

Artificial Intelligence Bot based on feedforward neural network for a Javascript (simpler) remake of the original The Binding of Isaac, JS, 2022

You can examine the demo from above picture HERE

How it works

The bot uses behavioral cloning technique in order to learn its parameters. That means that it trains on fixed collected data set. To obtain it recordings with game parameters and corresponding key presses of a real person playthrough are collected. Then this data is used to train a model.

The goal is to win against a certain TBoI boss named Duke of Flies I.

In this work the next 16 parameters were monitored during recordings:

1. hp - character's health points, 3 is max and 0 is game over
2. playerX - character's X coordinate, in pixels
3. playerY - character's Y coordinate, in pixels
4. bossDx - difference between boss's X coordinate and character's X coordinate, in pixels
5. bossDy - difference between boss's Y coordinate and character's Y coordinate, in pixels
6. bossD - euclidean distance between boss and character, in pixels
7. bossAngle - angle between boss and character, in radians
8. minNearDx - difference between nearest minion's X coordinate and character's X coordinate, in pixels
9. minNearDy - difference between nearest minion's Y coordinate and character's Y coordinate, in pixels
10. minNearD - euclidean distance between nearest minion and character, in pixels
11. minNearAngle - angle between nearest minion and character, in radians
12. enNearDx - difference between nearest enemy projectile's X coordinate and character's X coordinate, in pixels
13. enNearDy - difference between nearest enemy projectile's Y coordinate and character's Y coordinate, in pixels
14. enNearD - euclidean distance between nearest enemy projectile and character, in pixels
15. enNearAngle - angle between nearest enemy projectile and character, in radians
16. score - current score, which is calculated as 20 x bossDamage + 10 x kills + 5 x hits - bulletsUsed - 5 x damageTaken - timePassed / 100

Movements and fire buttons were also monitored.

After some matches (around 40-50) all recored data was downloaded with the help of written saveRecords method into json file and used for training.

For training the python's tensorflow library was used: .

A simple feedforward neural network with one hidden layer with 384 neurons and relu activational function was used. Network predicts 8 keypresses with 3 wait actions (for horizontal movement, for vertical movement and for firing).

After that model's weights was converted to a json file format and inserted into nn.js script. A feed forward prediction algorithm was written in pure JS.

The quality of the model was estimated via implemented win rate metric. By learning on 50 games, the win rate was around 40 percents.