Skip to: Site menu | Main content

Motivation: Main Research

As the observing technologies (sensors, cameras, camcorders, etc.) have matured over the years and the difficulties of data sharing have decreased, the amount of observation data has become immensely available (see this YouTube Video of swarm of starlings by National Geographic). My research is motivated by the need of making valid, useful and swift scientific discoveries from observation. For example, the story of how universal law of gravity was discovered traced back to Kepler's work on observation data of Mars' orbit.

To make a more precise statement of my research, the focus of my research is about designing and developing effective and efficient algorithms for making such discoveries. There are two major directions about my research:

In order to make accurate, convergent, and effective algorithms, I use theories and techniques from machine learning, numerical ODE/PDE, scientific computing, approximation theory, inverse problems, probability and statistics. Oftentimes, due to the gigantic size of the data set, the algorithms have to be scalable and efficient, hence I employ various techniques which reduce computing time: multi-scale representation, iterative solver, dimension reduction, domain decomposition, parallel computing, GPU computing (check out the Code section for details), etc. My research is interdisciplinary, it requries concrete knowledge of the observation technologies and underlying system, and it has applications in physics, biology, social science, and related engineering fields.

Motivation: Learning Dynamics

Dynamical systems have been used to model and describe complex physical, biological, social, and economical phenomena, see here. We are interested in developing and analyzing data-driven methods in deriving, validating, and improving complex dynamical system modeling. We are especially interested in a special kind of dynamical systems, namely the collective dynamics. It is also known as self-organized dynamics, as it displays self-organization in the system. The study of collective dynamics is driven by the need to understand some of the basic collective behaviors in bacterium, animals, human beings and even AI-controlled robots, such as flocking, milling, swarming, and rendezvous. Moreover, we are also interested in studying how emergent behaviors (clustering, flocking, milling, swarming, synchronization, etc.) appear in collective dynamics from observation data.

We developed a non-parametric learning approach to infer interaction laws (how agents interact with each other in the collective dynamics setting) in [1] as an extension of the paper, Inferring interaction rules from observations of evolutive systems I: the variational approach, by Maggioni, et al. We design and develop the learning program, and we show that it is well-posed, with excellent converngence properties as well as performance guarantees on how well our estimators are predicting the governing structure of the dynamical system, and the dynamical trajectories over the training time interval. We also tested several semi-supervised learning cases (model-selection) in order to showcase the excellent predication capability of our estimators.

In [2], we focus on expanding the learning approach proposed in [1] in three major directions: to study the large-time behavior of our estimators, to expand the learning of more complex dynamical systems (dynamics with 2-variable dependence interaction laws), and to discover hidden parametric stucture from the estimators.

In [3], we apply our method in learning gravitational dynamics to student some of the important historical astronomical events (discovery of precession of Mercury's orbit as well as general relativity effects on Mercury's orbit) from the current NASA JPL's Horizons data on our Solar system.

In [4], we provide convergence analysis of the expanded systems introduced in [2].

In [5], we extend our learning approach to collective dynamics on non-Euclidean manifolds, and investigate the convergence property of our estimators on these manifolds.

List of publications in learning dynamics:

  1. F. Lu, M. Zhong, S. Tang, M. Maggioni. Nonparametric inference of interaction laws in systems of agents from trajectory data (arXiv page), PNAS, 116 (3), 14424 - 14433, June 2019.
  2. M. Zhong, J. Miller, M. Maggioni. Data-driven Discovery of Emergent Behaviors in Collective Dynamics (arXiv page), Physica D: nonlinear phenomenon, 411 (?), 132542, October 2020.
  3. M. Maggioni, J. Miller, M. Zhong. Data-driven discovery of important historic astronomical events using the NASA JPL's Horizons data, In Preparation.
  4. M. Maggioni, J. Miller, S. Tang, M. Zhong. Learning interaction kernels in second order heterogeneous systems of agents from multiple trajectories, In Preparation.
  5. M. Maggioni, J. Miller, H. Qiu, M. Zhong. Learning Dynamics on non-Euclidean Manifolds, In preparation.

Examples -- Opinion Dynamics

The Opinion Dynamics (OD) studies how people's opinions interact with each other and how consensus is formed. It was first introduced as a discrete model by U. Krause in his paper, "a discrete nonlinear and non-autonomous model of consensus formation", in 2000. S. Motsch and E. Tadmor in their 2014 paper, "heterophilious dynamics enhances consensus", gave a very good summary of various kinds of opinion dynamics and extension.

Comparison of Interacton Kernels Comparison of Trajectories

Examples -- Lennard-Jones Dynamics

The Lennard-Jones Dynamics (LJ) uses an interaction law which is induced by a Lennard-Jones potential. It is used mostly in studying molecule formation and other molecular dynamics in quantum chemistry.

Comparison of Interacton Kernels Comparison of Trajectories

Examples -- Predator-Swarm Dynamics

The Predator-Preys Dynamics (PS) describes how a single predator interacts with a swarm of preys in a heterogeneous-agent system. We use the model instroduced by Y. Chen and T. Kolokolnikov in their 2013 paper, "a minimal model of predator-swarm interaction". We consider two different kinds of Predator-Swarm dynamics, Predator-Preys Dynamics of first order system (PS1), and Predator-Preys Dynamics of second order system (PS2).

Results from learning PS1.

Comparison of Interacton Kernels Comparison of Trajectories

Results from learning PS2.

Comparison of Interacton Kernels Comparison of Trajectories

Examples -- Phototaxis Dynamics

The Phototaxis Dynamics (PT) models how phototatic bacteria moves toward or away from stimulus of light. We use the model introduced by S. Ha and D. Levy in their 2009 paper, "particle, kinetic and fluid models for phototaxis".

Comparison of Interacton Kernels on v Comparison of Interacton Kernels on xi Comparison of Trajectories

Attention !!

The figures and movies for the examples of Cucker-Smale Dynamics, Self-Propelling Particle Dynamics, Synchronized Oscialltors Dynamics, Gravitational Dynamics will be made available after [3] is accepted and made publically available. If you are interested in those examples are used in [3], please check out the arXiv link of our paper.

Examples -- Cucker-Smale Dynamics

The Cucker-Smale Dynamics (CS) models how flocking is formed. F. Cucker and S. Smale in their 2007 papers, "on the mathematics of emergence" and "emergent behavior in flocks", introduced such elegant and simple alignment model to describe the flocking formation.

Examples -- Self-Propelling Particle Dynamics

The Self-Propelling Particle Dynamics covers a wide range of collective dynamics, which models particle dynamics where an individual particle can provide self-accelerating and self-decelearting force as well as collective interaction with other particles. We consider here two kinds of self-propelling particle dynamics: one is the fish milling model in 2D, the other one is the fish milling model in 3D within a fluid envrionment.

Fish milling model in 2D (FM2D) describes how these self-propelling particles form milling pattern in 2-dimensional space. We use the 2006 paper by M. R. D'Orsogna, Y. L. Chuang, A. L. Bertozzi, and L. S. Chayes, "self-propelled particles with soft-core interactions: patterns, stability, and collapse", and a follow-up paper in 2018, by Y. L. Chuang, M. R. D'Orsogna, D. Marthaler, A. L. Bertozzi, and L. S. Chayes, "state transitions and the continuum limit for a 2D interacting, self-propelling particle system".

Fish milling model in 3D (FM3D) within a fluid envrionment describes how self-propelling swimmers display corresponding emergent behavior (flocking or milling) in a fluid medium. We use the paper by Y. L. Chuang, T. Chou, M. R. D'Orsogna, "swarming in viscous fluids: three-dimensional patterns in swimmer- and force-induced flows", in 2016.

Examples -- Synchronized Oscillators Dynamics

The Synchronized Oscillators Dynamics models how a system of agents swarm and have their internal states locked in synchronization. We use the model proposed by K. P. O'Keeffe, H. Hong, and S. H. Strogatz, in their 2017 paper, "oscillators that sync and swarm". It is a combination of swarming in space and synchronization of phases or other internal states. For other purely synchronization models, please see S. H. Strogatz's review paper, "from Kuramoto to Crawford: exploring the onset of synchronization in populations of coupled oscillators", in 2000.

Examples -- Gravitational Dynamics in Our Solar System

The planetary movement of astronomical objects in our Solar system can be considered as a collective dynamics with a simplified Newton's gravitational law (see [3] for details). We consdier here the inner Solar system with 5 astronomical objects: Sun, Mercury, Venus, Earth and Mars. And we study the learnability of such interacting system without knowing the planetary orbits are elliptical or the form of the gravitatoinal force havin the famous 1/r^2. We are able to learn the masses of each astronomical object and the gravitational form all together through a delicate de-coupling procedure. And we are ready to apply our method to the JPL's Horizons data.