A neural network approach to microswimmer locomotion

Abstract

We propose a way of controlling microswimmer models with artificial neural networks. First, we review the physics of low-Reynolds-number swimming and the bead-arm model for microswimmers, and we give a brief account of Reinforcement Learning, artificial neural networks and the genetic NEAT learning algorithm. The discussed concepts are then put to use by training Reinforcement Learning agents to perform swimming motion. This is achieved by evolving mappings between a microswimmer's arm lengths and the forces on its arms. With this approach, we successfully train agents to perform locomotion on various swimmer types in one, two and three dimensions. Finally, we outline a method to perform steering in one dimension and train a steering agent to follow the gradient of a chemical field. We explore multiple ways an agent can measure the chemical field and show how the presence of noise in a temporally sensing agent can lead to the well-known run-and-tumble behavior of various microorganisms.