pAInball - AI Plays Space Cadet Pinball

Summary

This is a fun project that aims to train an Neural Network to play the famous 3D Pinball - Space Cadet, a game bundled with Windows 95 and XP.

To build this project, special thanks have to go to k4zmu2a/SpaceCadetPinball repo. Refer to that project for more information about porting Space Cadet.

How to build and play the game

You need to download the game resources (.dat & audio files) and place them in the bin/ directory. These resources are not included, but should be easy to find - just google it.

Afterwards, follow the below steps.

Compiling

On Windows

TBD - please contribute :)

On Linux

Install devel packages for SDL2, SDL2_mixer and SDL2_ttf.
Insatll cmake and ninja and compile with CMake; tested with GCC 10, Clang 11.\

At the root of the repo run:

cmake -GNinja .  
ninja

The executable should be under folder bin/SpaceCadetPinball. You can play it by running:

./bin/SpaceCadetPinball

On macOS

TBD - please contribute :)

Changes done to SpaceCadet Source code

In order to enable the AI to play the game, the source code of SpaceCadet had to be enhanced. These are the topline changes:

• Thread to output to stdout the inputs required for the Neuran Network
• Thread to read from stdin the outputs of the Neuran Network and to perform the commanded action
• By using the SDL2_ttf package, visualize on the screen the inputs to the Neural Network

See the game feedback

You can enable the game feedback via the stdout by enabling the "Debug Overlay" option:

AI that Sh*t 🤖

A Neuran Network has been developed to play the game. The NEAT genetic algorithm is used to create a generation, find the best performer(s), perform crossover and mutations and repeat. The best performer(s) of each epoch are stored, to allow continuing the training from a specific generation.

Signs of playing "intelligence" start to show up after ~100 generations, but your results may vary due to the random starting point.

Neural Network architecture

A Neural Network with 2 hidden layers of 7 nodes each was used.

Inputs:

• Location of the ball (X, Y)
• Speed of the ball
• Direction of the ball (X, Y)

Outputs:

• Do nothing
• Actuate Left Flipper
• Actuate Right Flipper

Parameters for the fitness function:

• Score
• Number of ball hits with the Left Flipper
• Number of ball hits with the Right Flipper
• Ball drained flag

Example:

Python Dependencies

See AIPlayer/pip_packages.txt

Install via: pip install -r pip_packages.txt

In order to tidy the multiple games on the screen to have a good overview, the system package wmctrl is used in run_game.py in function resize_window(). Install it based on your OS.

Run the AI

Execute the following to start from scratch (random Neural Networks):

python AIPlayer/run_game.py

Execute the following to start from specific parents

python AIPlayer/run_game.py <path_to_parent_0_weights.csv> <path_to_parent_1_weights.csv> ...

For the AI to receive game feedback, make sure you enable the Debug Overlay in the game settings:

AI Project structure

All the logic related to the AI is under AIPlayer.

• neural_ntw.py: Implementation of a Neural Network class
• nn_genetics.py: Implementation of the NEAT genetic algorithm for NN
• nn_visualization.py: Implementation of helper classes to visualize the Neural Networks and their weights (see Special Thanks)
• run_game.py: Main script

At the end of each epoch the script outputs the following:

• All parents in csv format at: AIPlayer/runs/<config_timestamp>/ep<X>_parent_<Y>_id<Z>_<fitness>.csv
• All parents visualizations in png format at: AIPlayer/runs/<config_timestamp>/ep<X>_parent_<Y>_id<Z>.png
• The last parents visualizations: AIPlayer/runs/<config_timestamp>/live_<Y>.png

AI Player Configuration

In run_game.py:

num_of_nns = 30 # Set the number of Neural Network Children per epoch
num_of_rand_children = 5 # Number for additional random Neural Networks per epoch
...
hidden_layers = [7,7] # Number of hiden layers and their nodes
start_timeout = 30; # Starting game duration per epoch. It is increasing by 10 sec every 10 epoch

Future Plans

• Train a Neural Network with more output actions
• Optimize fitness function to punish erratic behaviour
• Use PPO Algorithm for evolving the Neural Network

Special Thanks and Useful Sources

• Space Cadet port: k4zmu2a/SpaceCadetPinball
• Useful library for Neural Network Visualization: jzliu-100/visualize-neural-network
• Multilinear NN example: miloharper/multi-layer-neural-network
• Genetic Algorithm for optimizing a NN: luaffjk/ga-mlp