Neuroevolution in Unity

Maze Runners: Neuroevolution in Unity

50 AI agents learn to escape a maze from scratch, using a genetic algorithm and small neural networks. There’s no pathfinding and there are no scripted routes; every path is discovered by evolution.

I originally made this in an airplane ride, i achieved agents that learn how to find a goal, although they struggled a lot with walls. They were simple genetic code based algorithms though. 
Recently i updated it with some more functionality like a small neural net for agent, so they learn behaviors  (like not bouncing into walls) rather than a specific map.

How it works

• Each agent has a neural network that reads 8 wall-sensing rays and a line of sight to the goal, then decides how to turn.
• The best agents survive each generation. The rest are bred from them with crossover and mutation.

Problems solved

• Agents spinning in circles: fixed with a no-progress timeout, loop detection and novelty search.
• Dead ends that look close to the goal: fixed with line-of-sight rewards for seeing the goal and leaving spawn behind.
• Inconsistent results from the same brain: fixed with a deterministic, fixed-timestep simulation.
• Progress stalling: fixed with adaptive mutation and growing episode lengths.
  
  Features
  • 4 levels of increasing difficulty, plus a test arena.
  • Start, pause and restart, with simulation speed from 0.1× to 30×.
  • Replay any past generation’s best agents, with their paths drawn on screen.
  • Toggleable sensor and goal-ray visualisation.
  • Live stats.
  Built with Unity and C#.