Research

I work on the analysis, distributed control, and communication-topology design of large-scale networked cyber-physical systems. The common thread across my projects is co-design: rather than fixing a network and then designing a controller for it, I treat the controller and the interconnection topology as one design problem, and solve it with dissipativity theory, convex optimization, and — increasingly — learning.

• Research areas: cyber-physical systems; network systems; multi-agent systems; control and optimization; applied artificial intelligence.

• The ADIIS Lab pairs this theory with hardware. We run TurtleBot 4 and Arduino-based robotic testbeds for multi-agent persistent monitoring, and I mentor doctoral, master's, undergraduate, and high school researchers on them.

• Full publication list: Publications · Google Scholar · GitHub · CV

• Also available: the full research portfolio as an illustrated PDF — every project with its figures, in one document.

T1.1. Dissipativity-Based Control and Topology Co-Design for DC Microgrids — 2023-Present

• DC microgrids built from distributed generators must regulate bus voltages and share load current in proportion to capacity, but droop control couples these two objectives so that improving one degrades the other, and the communication topology that supports the distributed controller is normally fixed by hand.

• We model the microgrid as a networked system of distributed-generator subsystems, establish each subsystem's dissipativity, and co-design the distributed controller and communication topology through linear matrix inequalities. Later work handles constant-power loads and input saturation as sector-bounded nonlinearities.

• The result is droop-free control that achieves voltage regulation and proportional current sharing at the same time, verified on islanded microgrids with 4 generators and 4 lines and with 6 generators and 9 lines. The 2025 American Control Conference paper was one of five Best Student Paper Award finalists.

• People: Mohammad Javad Najafirad. Earlier course-project contributions from Akshitha Revu and Bhautik Patel.

• Papers: [C25] [C26] [C28] [C29] [A17] [A19] [A20] [A27]

T1.2. Dissipativity-Based Co-Design for Robust AC Microgrids — 2025-Present

• An AC microgrid has to solve three problems at once — regulate voltage, synchronize frequency, and share active and reactive power proportionally — and its dq-frame dynamics are nonlinear and strongly coupled, which is why most distributed AC designs treat the three objectives separately.

• We build a hierarchical architecture in which local controllers at each generator and a distributed global controller are linked through both the physical and the communication network, then co-design the global controller and that topology by the same dissipativity-based route as the DC case.

• The design holds all three objectives simultaneously and stays robust to disturbances, demonstrated on an AC microgrid with 4 distributed generators, 8 lines, and 2 loads. Presented at the 2026 European Control Conference in Reykjavik.

• People: Mohammad Javad Najafirad, with Lei Wu (Stevens) and Panos J. Antsaklis (Notre Dame).

• Papers: [C27] [A25] 

T1.3. Cooperative Frequency Support Control for Offshore Wind Farms — 2025-Present

• As wind generation displaces synchronous machines, system inertia falls and wind farms are increasingly required to supply primary frequency support themselves — but turbines within a farm sit at different operating points and can spare very different amounts of power.

• We developed a two-level semi-distributed cooperative control strategy: a farm-level model-predictive controller derives the optimal frequency trajectory and the resulting power burden, turbine-level controllers share that burden cooperatively according to each turbine's extractable capacity, and an explicit time coordination scheme links the two levels.

• Against fixed-droop and variable-droop baselines, the strategy improves the grid frequency nadir while leaving turbines with more remaining extractable capacity, and it holds up under varying wind speeds.

• People: Zhuzhu Wang, Yikui Liu, and Lei Wu (Stevens).

• Papers: A journal manuscript is under review.  

T2.1. Hierarchical Control and Topology Co-Design for Robust Networked Systems — 2024-Present

• Synthesizing a controller for a large networked system centrally does not scale, and it destroys the property you most want to keep: that adding or removing a subsystem should not force a redesign of everything else.

• We enforce dissipativity locally at each subsystem, then use those certificates to co-design the distributed controller and topology through linear matrix inequalities. We then extended this to the data-driven setting, with only input-state-output trajectories available, via the matrix S-lemma.

• Model-based and data-driven variants carry the same robustness guarantees, demonstrated on DC microgrid examples. A journal version is in preparation.

• People: Zihao Song, with Hai Lin and Panos J. Antsaklis (Notre Dame).

• Papers: [A28]  

T2.2. Graph Neural Network-Based Distributed Control — 2024-Present

• Distributed optimal control laws for networked systems are expensive to compute and have to be recomputed whenever the network changes, and the graph-neural-network controllers proposed to sidestep that are trained offline and centrally, which gives back the scalability they were meant to buy.

• We put a graph recurrent neural network controller directly in the feedback loop and train it online and in a distributed manner, so each node learns using only information from its neighbors.

• The learned distributed controller tracks the performance of a centralized LQR solution — the global optimum — in both average states and average control inputs.

• People: Zihao Song, with Hai Lin and Panos J. Antsaklis. Related course project on recurrent-equilibrium-network feedback controllers by Jiafeng Yang.

• Papers: [A22] 

T2.3. Funnel-Based Online Recovery Control — 2025-Present

• Offline co-design assumes the system you designed for is the system you have. When a subsystem is damaged or its dynamics drift, the guarantees lapse, and for nonlinear systems with unknown dynamics there is no model to fall back on.

• We learn the unknown dynamics with a recurrent-equilibrium-network model inside a virtual system, then drive the actual system to follow the virtual trajectory through a sequence of waypoints, each surrounded by an invariant funnel that the state is guaranteed not to leave.

• This gives an online resilience layer that complements the offline co-design, demonstrated by recovering a DC microgrid in real time.

• People: Zihao Song, with Hai Lin and Panos J. Antsaklis.

• Papers: [A24]

T2.4. Control and Topology Co-Design for Vehicular Platoons and Rigid-Body Networks — 2021-Present

• A platoon must stay string stable — disturbances must not amplify down the line — as vehicles join and leave; the same question governs formations of under-actuated rigid bodies such as quadrotors. Backstepping becomes intractable as formations grow, disturbances are mismatched, and topology affects stability.

• We co-design the distributed controllers and communication topology under a disturbance string stability constraint, decentralized by a Sylvester's-criterion-inspired argument; mismatched disturbances get adaptive backstepping on vector string Lyapunov functions, and mesh stability extends this to rigid bodies.

• Merging and splitting are then handled by compositional gain updates rather than full re-synthesis, and the analysis identifies back-to-front communication links as the most important for string stability. The rigid-body extension gives quadrotor formations with mesh stability guarantees under merging and splitting.

• People: Zihao Song, with Hai Lin and Panos J. Antsaklis.

• Papers: [J14] [J15] [C21] [C24] [A11] [A15] [A21]. 

T3.1. Hierarchical Dissipativity-Based Analysis and Control of Spreading — 2025-Present

• Spreading processes are usually analysed at a single level, either node by node or through a network-wide aggregate, which either misses group structure or hides the individual dynamics that drive it.

• We model the spreading network as a hierarchical system of nodes, groups, and the whole network, establish dissipativity of the nonlinear node dynamics, and generalize our topology design to that setting. Analysis and control become one convex linear matrix inequality problem per level, with no iterative optimization.

• The resulting controllers guarantee stability, dissipativity, and mesh stability of the spreading network, and therefore robustness to disturbances. Numerical studies show better infection containment, better control efficiency, and better disturbance robustness than conventional methods.

• Papers: [C31] [A23]; a journal version is in preparation. This direction is the basis of an NSF Engineering Research Initiation proposal.

T4.1. Inventory Consensus and Topology Co-Design for Supply Chain Networks — 2023-Present

• Supply chain networks have to hold inventory near target levels across suppliers, distributors, and retailers while demand varies and while transport delays, waste, and decay act on every link.

• We formulate inventory consensus across the network as a networked system control problem, with each facility a subsystem and each ordering and delivery relationship an interconnection, then apply dissipativity-based distributed control and communication topology co-design to regulate inventory levels.

• The co-designed topologies improve consensus performance, measured by percentage mean absolute error, against both reference and fully connected topologies, on simulated customer demand data. Inventory tracking is the next extension and is being formulated now.

• People: Jessica Joseph contributed the sourcing and preprocessing of demand, supply, and inventory datasets.

• Papers: [C30] [A18]. 

T5.1. Energy-Aware Persistent Monitoring on Robotic Testbeds — 2024-Present

• Persistent monitoring policies are designed as though agents run forever, but a real robot has a battery, and no monitoring decision can account for energy without a usable model of how motion depletes charge.

• We characterized TurtleBot 4 battery depletion under different motion profiles to quantify performance-energy trade-offs, and built a physics- and data-informed model linking PWM duty cycle, operating time, and state of charge for a simulated Arduino robot, using sparse identification of nonlinear dynamics.

• The trade-off characterization supports energy-aware monitoring decisions on the testbed, and the state-of-charge models predict depletion closely enough over a 300-second horizon to be used for planning. TingYou Chung defended his M.S. thesis on this work in April 2026.

• People: TingYou Chung, Vidyut Pradeep, Siyao Wu, Travis Chi, Gabriel Fanshteyn, Abhishek Prithvi Teja Angadala.

• Papers: [A26]; M.S. thesis, "Performance-Energy Trade-offs in Energy-Aware Persistent Monitoring Systems" (TingYou Chung, 2026).

T5.2. Actively Collaborative Multi-Agent Persistent Monitoring — 2024-Present

• Standard multi-agent monitoring is passive divide-and-conquer: the targets are partitioned, each agent takes a share, and the agents never help one another even when one is overloaded and another is idle.

• We are developing a formulation in which agents collaborate actively and the cost of that collaboration is carried explicitly in the objective, so cooperation happens when it pays for itself and not otherwise. The formulation has now moved into implementation in the lab's simulation environment.

• Results so far have been presented as student posters at the ECE Research Expo and the 3rd iCNS Symposium. A paper is in preparation.

• People: William Getts.

T5.3. Safety-Aware and Learning-Based Multi-Agent Control — 2025–present

• Learned multi-agent policies scale where hand-designed ones do not, but they come with no safety guarantee, and the usual fix — a central monitor that vetoes unsafe actions — gives up the decentralization that made learning attractive.

• We developed a fully decentralized, safety-aware multi-agent deep reinforcement learning algorithm that combines a graph encoder, a node-uncertainty state estimator, and a discrete-time control-barrier-function safety layer applied directly to an actor-critic policy, so every agent enforces its own safety constraint from local information.

• On custom simulation environments — one with 50 nodes and eight targets, a second with 100 nodes and 17 — the algorithm shows clear performance gains over baseline decentralized policies. The paper was accepted to the IEEE-HKN "Innovating the Future" papers conference.

• People: Theodore Rogalski (undergraduate thesis, defended April 2026; M.S. thesis student from Fall 2026) and Vedaang Verma (control-barrier-function safety layer).

• Papers: [C32]

Thrust 1 — Decentralized Synthesis of Distributed Controllers for Networked Systems — 2021-Present

• Analysis and controller synthesis for a networked system with arbitrary interconnections is normally a centralized computation, which does not scale and has to be redone whenever the network changes.

• We introduced a class of matrices we call network matrices and exploited Sylvester's criterion to turn the underlying linear matrix inequality problems into iterative processes that each subsystem can execute using only local and neighbour information.

• This yields decentralized dissipativity analysis and dissipativating controller synthesis for networked systems with arbitrary interconnections. The MED 2022 paper received the conference's Best Paper Award.

• Papers: [BC1] [J13] [C18] [A10]

• Code: github.com/shiran27/NetworkedSystemAnalysisAndControl 

Thrust 2 — Analysis and Topology Synthesis for Non-Linear Networked Systems — 2021-2024

• Large-scale networked systems built from interconnected non-linear subsystems appear throughout engineering, but the analysis and topology-design tools available for them lag well behind the linear case.

• We developed centralized and decentralized — and compositional — techniques for both analysis and topology synthesis, exploiting the dissipativity properties of the subsystems to formulate linear matrix inequality problems that enforce stability or a desired dissipativity property on the network as a whole.

• The same machinery designs the physical interconnection topology and the information-flow topology, and it underpins the microgrid, supply chain, and spreading work.

• Papers: [C20] [C22] [A12] [A13]

Thrust 3 — Real-Time Estimation of Passivity Indices — 2021-2023

• Passivity indices are useful for analysis, controller synthesis, and fault tolerance alike, but they are normally computed from a model — which is exactly what you lack in the situations where they would help most.

• We studied a class of fractional function optimization problems and exploited their structural properties to derive an efficient online solution, then applied it to estimate the passivity indices of a dynamical system from its input-output data alone.

• The estimator runs online and tracks stability and passivity metrics as the system operates.

• Papers: [C19] [A9]

Thrust 4 — Symbolic Control Problems via Signal Temporal Logic — 2021-2023

• Autonomous systems often have to satisfy several spatio-temporal specifications at once. Signal temporal logic expresses these cleanly and turns the control problem into a symbolic one, but the robustness measures used to optimize against STL specifications are non-smooth, which blocks gradient-based solvers.

• We developed new smooth STL robustness measures and derived expressions for both their approximation error bounds and their explicit gradients.

• Symbolic control problems can then be solved efficiently and effectively by gradient-based optimization, with a quantified bound on what the smoothing costs.

• Papers: [A14]

• Code: github.com/shiran27/Symbolic-Control-via-STL

Thrust 5 — Robust Approximate Simulation of Piecewise Affine Systems — 2021-2022

• Approximate simulation relations let you design a controller on a low-dimensional abstraction and carry the guarantee back to the full system, but no such strategy existed for piecewise affine systems.

• Co-supervising Zihao Song, then a third-year graduate student, we developed one for the first time, under bounded disturbances.

• Papers: [C17] [A8]

Thrust 1 — Multi-Agent Persistent Monitoring on Networks — 2019-2023

• Persistent monitoring asks a team of agents to keep a set of targets under observation indefinitely, with the agents' trajectories as the decision variables — a class that covers a wide range of surveillance, inspection, and data-collection problems.

• We designed a distributed online solution based on receding horizon control and a centralized offline alternative based on greedy cycle search and perturbation analysis. Both rested on explicit solutions we derived for a class of rational function optimization problems and a novel "auxiliary target" construction.

• The two solutions bracket the problem from opposite directions — online and distributed against offline and centralized — and the minimax variant handles worst-case rather than average uncertainty.

• Papers: [J8] [J9] [J11] [C13] [C14] [C16] [A2] [A4] [A6] [A7]

• Code: github.com/shiran27/PersistentMonitoring · Simulator: bu.edu/codes/simulations/shiran27/PersistentMonitoring

Thrust 2 — Multi-Agent Distributed Estimation on Networks — 2020-2025

• When several mobile sensors monitor a collection of targets, each target's estimation error covariance grows while unobserved and shrinks while observed, so the sensors' routing problem is really a problem about covariance dynamics.

• We derived explicit expressions for the error covariance dynamics under different monitoring settings and used them to design the sensors' controllers, giving a receding-horizon-control and machine-learning based distributed online solution, alongside a minimax sensor allocation solution computed centrally offline.

• Later work at Stevens revisited the nonlinear covariance dynamics with neural-ODE and deep-network approximators.

• People: later approximator work by Alexander Zhavoronok and Dominic Magats.

• Papers: [J10] [C15] [A5]

• Code: github.com/shiran27/PMG_Error_Covariance

Thrust 3 — Improved Performance Bounds and Multi-Agent Coverage Control — 2018-2021

• Many resource allocation problems reduce to maximizing a submodular set function, where the greedy solution comes with the classical performance bound — a bound that is often loose and, when tightened, expensive to evaluate.

• We derived a new performance bound from extended greedy iterations that characterize the objective's curvature at almost no extra cost, then combined greedy initialization with projected gradient ascent to solve coverage without a mixed integer program, and solved heterogeneous team composition jointly with coverage.

• Empirically, the new bound is much tighter than the classical one and is essentially free to compute.

• Papers: [J6] [J12] [C23] [A3] [A16]

• Code: github.com/shiran27/CoverageControl · github.com/shiran27/OptimalTeamComposition 

• Simulators: bu.edu/codes/simulations/shiran27/CoverageFinal · bu.edu/codes/simulations/shiran27/OptimalTeamComposition

Thrust 4 — Overcoming Local Optima in Distributed Multi-Agent Optimization — 2017-2019

• Most multi-agent control and optimization problems of interest are non-convex, so gradient-based distributed methods converge — but frequently to a local optimum that can be arbitrarily worse than the global one.

• We built a systematic distributed online algorithm around the boosting functions idea, with convergence guaranteed through a novel step size adaptation law, and selected effective boosting functions by exploiting the structural properties of the objective.

• The algorithm escapes local optima with a convergence guarantee intact, and solved both a class of coverage problems and a class of consensus problems.

• Papers: [J7] [C12] [A1]

• Code: github.com/shiran27/CoverageControl · Simulator: bu.edu/codes/simulations/shiran27/CoverageFinal

Thrust 1 — Real-Time Non-Intrusive Load Monitoring — 2015-2017

• Non-intrusive load monitoring infers which appliances are running in a household from the aggregate power signal at a single meter, which makes it a cheap route to the data that power network planning and operational management need — and a natural fit for the smart grid.

• We developed a method that decomposes the residential power signal into orthogonal components using the Karhunen-Loève transform, then biases the appliance prior probabilities using appliance usage patterns and a fuzzy system, and extended it to operate under residential solar power influx and supply voltage fluctuations.

• The method also produces short-term load forecasts, and we implemented it in real-time hardware. It received three President's Awards for Scientific Research from the National Research Council of Sri Lanka, a patent, and a 2025 Certificate of Commendation from the National Science Foundation of Sri Lanka.

• Papers: [J3] [J4] [J5] [C6] [C7] [C11] 

Thrust 2 — Complete Reconnaissance Robotic System — 2014-2015

• A reconnaissance robot is only useful if it can localize itself, map its environment, plan and accurately track a trajectory, and record sensory data — all of it on board, and all of it at once.

• For an omnidirectional robot platform, we designed a sensor-fusion-based localization system, a stereo-camera-based mapping system, and a non-linear dynamics-based trajectory tracking controller.

• We implemented these together on an on-board real-time embedded system built from six Arduino microcontrollers and a pcDuino on-board computer.

• Papers: [J1] [C1] [C2] [C3] [C4] [C5]