Unity Machine Learning Library

Goals of this project are:

• implement machine learning algorithms from scratch;
• create visualisations for these algorithms;
• use the algorithms to solve game development related problems;
• re-learn calculus :)

Long term goal is to create a "plug-and-play" AI for racing games, which is agnostic, as much as possible, of the game.

Install&Usage

This project requires Unity 6, however, the core library can be used in earlier versions.

Samples

Multi Arm Bandit

Implements and visualises the multi armed bandit problem

Q Learning

Implements and visualises Q learning algoritm applied to finding an optimal path in a 1D and 2D grid world.

Self Driving Cars (Reinforced Learning and Genetic Algorithms)

Implements feed forward neural networks and reinforced learning to have cars learn to drive any random track.

Since optimisation/back-propagation is not yet implemented, the simulation manager uses a genetic algorithm-like approach to select the best cars.

The simulation might take some time to find a neural network (weights/biases) that can drive around the track.

Use WASD and right-mouse button to navigate and look around. Use left-mouse click to speed up movement (FreeCamera script).

An inspector is available in Tools > Piron Games > MLLib > Self Driving Cars > Car Inspector. Click on a car to inspect its parameters (currently lacks neural network visualisation).

TODO

Core library:

• implement an optimization algorithms (gradient descent)
• implement neural network back-propagation
• implement DQN
• documentation
• unit tests

Samples/Visualisation:

RL self driving cars:

• implement neural network visualisation in Car Inspector tool using GraphView
• make Car Inspector available for Canvas as well
• improved ability to manipulate neural network snapshots for the self driving cars simulation (copy/paste from one car to another, etc)
• experiment with short and long range sensors for the cars
• improve the simulation manager spawning (loitering -> from first waypoint, out of bounds -> from last waypoint)
• further experimentation for the track learning manager using genetic algorithms

License

This project is licensed under MIT license.

Uses assets from Kenney, which are licensed under Creative Commons Zero.

Uses Unity Vehicle Tools, which has no license.