Description

The trex runner game is from wayou repo's in here. What I do here is just adding the setup for reinforcement learning.

This is my first time to build a reinforce learning project. Hope I can improve more, feel free to reproduce or modify the code.

Author

Rakka Alhazimi, October 2021

Requirements

First, let me tell you the summary of what I was use here

Software:

Python v.3.7.7
OS : Windows 10

Python side-libraries:

Selenium
Keyboard
Tensorflow
Numpy

Web-Driver for Selenium:

Chrome

Any version is okay I guess, as long as the dependencies are fulfilled.

Installation

To make this project runs well on your machine, what you need to do is:

Install Python

After you install python, I suggest you to make an venv with command:
```
python -m venv [name]
```
it'll build a folder in your directory right away, and to activate the env use:
```
[name]\Scripts\activate.bat
```
Then install all the dependencies using:
```
pip install selenium, keyboard, tensorflow, numpy
```
You can jump right into installing the dependencies if you already have python and venv intact.
Install Chrome Web-browser

Chrome lets us to access console.log message from selenium. Therefore, Google Chrome is a must.
Download webdriver for chrome here

Match the driver with your Chrome version, after you download it, move the driver in your current directory (repo directory / same level with main.py).
Start your local server to serve index.html
You can use live server from VS-Code or anything else. Then copy your localhost link like http://127.0.0.1:5500/ to the CONFIG variable in utils/config.py file.
Start the program The main file is main.py, run this file to start the program.
```
python main.py
```

Configuration

You can update or change the settings and hyperparameter in the utils/config.py file.
Hyperparameter that you must know:

SERVER: Your local server link to be redirected by the webdriver
TRAIN : Whether or not you want to train the model
CONTINUE: Whether or not you want to continue from the model checkpoints
EPISODES: The number of episodes you want to run

Training Note:

How do I train this dino agent:

Go for 100 - 500 Episodes with 1e-2 Learning Rate
Set the max reward = 500.000

Model Used

Regular policy gradient method

Main Idea

I try to make a reinforcement learning project with the in-game parameters. Other peoples are likely to snipe the screen pixel or taking multiple screenshot to gain the parameters of the state. Therefore, I would like to try different approach by peeking the game-code and get its variable. It's much faster and memory-friendly. The rest is still the same, the approach, model and etc.

I hope you can enjoy training your agent using the default model or you can also build your owns model to beat this game, any modification is welcomed. You can also change how the parameters is recorded inside the src/js/watcher.js file.

rakkaalhazimi/tRex-reinforcement-learning