Q-learning for real time control of heterogeneous microagent collectives

Ana Rubio Denniss, Laia Freixas Mateu, Thomas Gorochowski, Sabine Hauert

Keywords: q-learning, reinforcement learning, microsystems, closed-loop control, optical control

Abstract: The effective control of microscopic collectives has many promising applications, from environmental remediation to targeted drug delivery. A key challenge is understanding how to control these agents given their limited programmability, and in many cases heterogeneous dynamics. The ability to learn control strategies in real time could allow for the application of robotics solutions to drive collective behaviours towards desired outcomes. Here, we demonstrate Q-learning on the closed-loop Dynamic Optical Micro-Environment (DOME) platform to control the motion of light-responsive Volvox agents. The results show that Q-learning is efficient in autonomously learning how to reduce the speed of agents on an individual basis.

One-sentence Summary: Q-learning is used to learn optimal strategies for optically controlling a collection of microscopic biological agents