Skip to content

Latest commit

 

History

History
98 lines (71 loc) · 5.1 KB

README.md

File metadata and controls

98 lines (71 loc) · 5.1 KB
Raw

N-ZeroX

(Neural)-Zero X is an AI that can play F-Zero X on BizHawk emulator using real time CNNs. This repository is part of a Universidad de Sevilla's final degree project.

Set-up environtment

To run this project, you need Python 3 and Bizhawk emulator.

Install 64-bit Python 3

This project was written for Python 3.7. Tensorflow requires 64-bit Python.

Install Python Dependencies

The following Python dependencies need to be installed.

  • Tensorflow 2.2.0
  • Keras 2.3.1
  • Pillow
  • matplotlib
  • mkdir_p
  • h5py

(Optional) Install CUDA and cuDNN

Although you can run Tensorflow on CPU, I'll recommend you to download and install TensorFlow GPU dependencies too.

TensorFlow Python cuDNN CUDA
2.2.0 3.5 to 3.8 7.6 10.1

Get BizHawk emulator

This project contains LUA script files ready to run on BizHawk emulator (tested on version 2.6.2). To get BizHawk you first need to install the prerequisites. Then you can download BizHawk and unzip it to any directory.

You will also need a F-Zero X ROM to run on BizHawk emulator.

Download Pre-trained Weights and Recordings

Download this data to run the demo. You can also download my recordings to train the models by yourself. These should be unzipped into the folder of this repository.

  • Save States - LUA scripts will access the saved states on states/[file].state.
  • Weights - Python scripts will access the trained models on weights/[model].hdf5
  • Recordings (Optional) - The recordings should be accessible as recordings/[recording]/[frame].png.

Recordings file is almost 1GB. It contains >12000 game samples as screenshots.

Usage Instructions

You must run .py files directly from console and .lua files from BizHaw Lua Console. You can find all lua files on scripts folder.

Running the Demo

These instructions can be used to run a demo on Mute City using the demo model.

  1. Download the save states and pre-trained model.
  2. Run predict-server.py using Python - this starts a server on port 36296 which actually runs the model.
    • You must specify the model you want to run. Use demo as first parameter.
    • You can pass a --cpu to force Tensorflow to run on the CPU.
  3. Open BizHawk and load a F-Zero X ROM.
  4. Turn off messages (View > Display Messages).
    • You don't have to do this, but they get in the way.
  5. Open the BizHawk Lua console (Tools > Lua Console).
  6. Load Demo.lua

This should automatically play Mute City time attack race. You can hit the arrow keys to manually steer the Blue Falcon. This can be used to demonstrate the AI's stability.

Generate your own training data

The first thing you need to train your model is training data. You can generate training data using Record.lua and RecordImput.lua.

  1. Open BizHawk and load a F-Zero X ROM.
  2. Open the BizHawk Lua console (Tools > Lua Console).
  3. Load RecordImput.lua
  4. Play for a while (I'll recommend to use a joystick or game controller)

A new folder will be created on recordings. A screenshot will be stored everytime you move the joystick playing the game with the imput value stored on steering.txt

Training the Model on Recordings

Once you have generated new recording, you probably want to try retraining the weights based off your recordings. To train a new model, run train.py [model]. You can also use --cpu to force it to use the CPU. Your trained model will be stored on weights/[model].hdf5

You can include train data from train.zip along your own recordings.

Play your trained model on a race

You can load the race savestate from states/MuteCityGP.state to test your new trained model. Remember to launch predict-server.py first and load Play.lua from Lua console.

Train AI to play on another track

You can use Record.lua and RecordImput.lua to generate training data for another track. Even for another game! Remember to use a different name as parameter when you train your model with train.py.

Reference Projects

  • NeuralKart - This project was forked from rameshvarun real time Mario Kart AI.
  • TensorKart - The first MarioKart deep learning project, used for reference.