Playing the game of snake with limited knowledge: Unsupervised neuro-controllers trained using particle swarm optimization

Methods in the domain of artificial intelligence (AI) have been applied to develop agents capable of playing a variety of games. The single-player variant of Snake is a well-known and popular video game that requires a player to navigate a line-based representation of a snake through a two-dimensional playing area, while avoiding collisions with the walls of the playing area and the body of the snake itself. A score and the snake length are increased whenever the snake is moved through items representing food. The game thus becomes more challenging as the score increases. The application of AI techniques to playing the game of Snake has not been very well explored. This paper proposes a novel technique that uses particle swarm optimization for the unsupervised training of neuro-controllers in order to play the game of Snake. The proposed technique assumes nothing about effective game playing strategies, and thus works with limited knowledge. Sensory input is also minimal. Due to the lack of similar AI-based approaches for playing Snake, the proposed technique is empirically compared against three hand-designed Snake playing agents in terms of several performance measures. The performance of the proposed technique demonstrates the feasibility of the approach, and suggests that future research into AI-based controllers for Snake will be fruitful.