Contributed talk

Controller Noise and the Evolution of Exploration and Exploitation

Vincent Ragusa, Jory Schossau, Arend Hintze

Noise in machine learning is often seen as the adversary: Model training methods are often born from trying to increase model robustness and generalization performance. While we try to remove the noise from our learning systems, natural brains perform admirably in noisy environments while being themselves composed of very noisy components (neurons and other processes). Here we explore how applying noise to an evolved agent's action-perception loop (sensors, hidden states, and outputs) affects their ability to solve an exploration and navigation task. We find that the application of low to moderate noise greatly improves the evolved agents. We also find that the benefit of noise decreases as the task biases toward exploration.