AI learns to play Snake!

A neural network learns to play snakes

Built with Rust and Macroquad game engine

Explaination & Timelapse Video

Controls

Tab - Enable/Disable visualization
Space - Slow down the simulation

Glossary

Neuro-evolution: A subfield of artificial intelligence and evolutionary computation that uses evolutionary algorithms to evolve artificial neural networks.
Genetic Algorithm (GA): An optimization algorithm inspired by the process of natural selection and genetics, used to evolve populations of solutions to optimization problems
Population: A collection of individuals (neural networks) that are subject to evolution in a neuro-evolutionary algorithm.
Island Model: A technique in neuro-evolution where multiple separate populations (islands, in code this is called streams) of individuals evolve independently, periodically exchanging individuals between islands to maintain diversity and promote exploration.
Mutation: The process of introducing random changes to the genetic material (weights, biases, or network structure) of individuals in the population. Mutation helps introduce new variations and explore the solution space

Overview

Setup

Every snake has a neural network that acts as its brain.
The snake can see in 4 direction. It can detect food, wall and itself in these 4 directions. Total number of inputs = 4 * 3 = 12
These 12 values are fed as an input to the neural network. The neural network then generates 4 values that indicate the threshold for actions - left, right, bottom and top.
Every generation has 5 streams (islands) of 1000 snakes each. The snakes in each stream evolve independently of the snakes from other streams
Occasionally best performing snakes from one stream are injected into another. This technique is called "Island Rejuvenation".

Algorithm

The simulation begins at Generation 0 with 5 streams of games, the individuals in each of these streams have randomly generated neural networks.
Each step, we update every game i.e pass the vision inputs to the neural network and have it decide on an action to take.
When the snake collides the walls or when it collides with itself, the game is flagged as completed.
Aditionally, games are marked complete when the snake isn't able to eat food for a certain number of steps. This is to prevent snakes from performing looping actions indefinitely.
We update each game in a generation until all the games are complete.
At the end of each generation, each snake in a stream is ranked based on how it performed.
Based on this ranking, parents are chosen for the next batch of snakes. Snake at rank 1 is more probable to be a parent compared to snake at rank 10.
Here's an example of how the population is distributed throughout every generation:
1. Top 1% of the snakes are retained for next generation
2. 50% of the population is newly generated using 2 snakes from the previous generation as their parents
3. 20% of the snakes are freshly generated with no connection to the past generations
4. The rest 29% of the population are all mutations of the current best performing snakes
Once we have a new population, we start a new generation. And the above steps are performed until the simulation is closed manually.
The above steps result in the snakes fine tuning their strategies which inturn lead to longer snakes.

Usage

Clone the repo

git clone git@github.com:bones-ai/rust-snake-ai.git
cd rust-snake-ai

Run the simulation

cargo run --release

Configurations

The project config file is located at src/configs.rs
Disable VIZ_DARK_THEME changes the theming
The streams feature is still experimental. A single stream with 1000 snakes will yield quick results.