Talk
Speaker: NAMAN KRISHNA PANDE, Indian Institute of Technology Ropar
Host: S Baskar
Title: Scientific Machine Learning for Complex Systems
Time, day and date: 11:30:00 AM – 12:30:00 PM, Monday, August 10
Venue: Online (https://meet.google.com/jxd-pimp-pox)
Abstract: Multi-agent systems (MAS) are ubiquitous to various natural phenomena such as flocks of birds, crowd motion, vehicular flow and insect swarms, etc. These systems involve a large number of interacting agents whose collective behaviour gives rise to complex emergent phenomena. Analysing and controlling such systems poses major problems in various scientific and engineering fields. Optimal control (OC) of MAS aims to learn strategies that enable the agents to achieve common or individual objectives while adhering to some dynamical constraints. We aim to develop and analyse these systems through the mean-field game theoretical approach that considers each agent as rational and utility-optimising and comprehends their dynamics at the macroscopic scale. Furthermore, we will utilise physics-informed neural networks (PINN) based models to learn the dynamics of such complex physical systems typically governed by a system of partial differential equations. However, despite enjoying great success, PINNs have some fundamental flaws when dealing with various complex physical problems. We aim to address various limitations of PINNs by developing robust and scalable physics-informed learning frameworks that incorporate improved neural network architectures, efficient sampling techniques and integration of discrete representations of the governing physics directly into the network architecture and training process.