Neural network learns to evade and capture

July 9, 2018

Watch these objects as they try and eat and avoid being eaten on a circular food-chain.

Three objects have a "rock-paper-scissors" relationship. Red eats green, blue eats red, and green eats blue. The red, green and blue objects co-evolve their evasive and predatory behavior.

While the system is symmetric, the performance of the three colors diverges as they fill different evolutionary niches. In this case, the red and green seem to have figured out a way to win out at the expense of the blue.

Their behavior is controlled by a neural net with a single hidden layer, and they are given a slightly delayed reflex with regard to information about their opponent's positions and velocities.

They are also allowed to move erratically and spontaneously through the use of an input that switches between positive and negative randomly with a mean frequency that is controlled though an evolutionary algorithm.

I will be giving a talk in Oxford this evening (9 July 2018) on this and other projects.

← Procedural Audio in Unity - Noise and Tone Virtual Reality Reading UIs →