Neural Networks for Complex Systems: From Epidemic Modeling to Swarm Robotics

Abstract

Machine learning refers to data-driven approaches first introduced from the 1950s, drawing inspiration from developments in neurosciences. An artificial neuron is the smallest computational unit in machine learning. Neural networks are obtained by combining multiple neurons into layers that progressively process input information to determine an output. Neural networks are trained to find complex relationships in large amounts of data, and over the last decades, they have enhanced several fields of research including physics. In this thesis, I work at the intersection of complex systems and neural networks. In the first part, I focus on applications in which neural networks are a tool for analyzing data generated by complex systems. In Paper I, I employ neural network in agent-based simulations of epidemics. In Paper II, I look at one of the main issues in real-world applications, handling incomplete datasets. In the second part, I shift to use physical systems to build neural networks. I employ robots to play the role of artificial neurons within a swarm. In Paper III, I introduce physical constraints and use movement to restructure the neural network. In Paper IV, I use a robotic swarm to realize a generative machine learning model that reconstructs light patterns. These robotic experiments can be seen as an intermediate step between artificial and biological neural networks, still keeping the programmable abilities of the former, but including some of the physical constraints of the latter.