Repo: Github – SnakeGameAI
I built a simple snake game and trained an AI to play it using NEAT (NeuroEvolution of Augmenting Topologies) machine learning. Demo below:
Overview#
The agent operates on a deliberately minimal input configuration: binary indicators for obstacles directly ahead and to either side, and the angle to the food. More complex state representations were also explored, but they tend to lead to more ineffective training.
Despite the limited information, the agent learns competent behaviour. It can navigate around the map well, avoiding additional obstacles I added, as well as its own body, especially during early and mid-game. However, as the snake has no global knowledge, it may trap itself in dead ends, which is what happened in the demo video as well. This becomes increasingly likely as the snake grows longer.
This project primarily served as an exploration of the strengths and limitations of NEAT. In hindsight, a classical path-finding algorithm like depth-first search, extended to consider its own body and its future position, would most likely outperform this and be simpler to implement. Nevertheless, this project was a fun investigation of NEAT, and highlights the importance of finding an appropriate approach for a given problem.