Cemal Tugrul Yilmaz

I am a PhD candidate in the Department of Mechanical and Aerospace Engineering at the University of California, San Diego under the supervision of Prof. Miroslav Krstic and Prof. Mamadou Diagne.

My research interests include model-free optimization, adaptive control, disturbance estimation and rejection, and the control of distributed parameter systems.

Education

University of California San Diego, Mechanical Engineering, PhD, 2020 - Present
Bogazici University, Mechanical Engineering, MS, 2015 - 2018
Bogazici University, Mechanical Engineering, BS, 2010 - 2015

Awards

Recipient of the 2023-2024 UC San Diego MAE PhD Outstanding Student of the Year Award (2024)
Awarded UC San Diego MAE First Year Fellowship (2020-2021)
Awarded Undergraduate Research Student Grant from Turkish Scientific and Technological Research Council (2014-2015)
Ranked in the top 0.05% in the National University Entrance Exam out of more than 1.5 million candidates (2010)

Projects

Accelerated and Unbiased Model-Free Optimization

Aim is to seek the unknown optimum θ* of the unknown cost function h(θ) without any gradient information.
Classical algorithm explores the search region through periodic oscillations, estimates the gradient and provides a biased estimate of the optimum due to persistent oscillation.
We present the first results achieving unbiased convergence to the unknown optimum θ*
- 1. at a user-assigned asymptotic or exponential rate [CDC23]
  2. more powerfully, in a user-prescribed time (for instance, in 2 hours, 10 seconds, or 1 millisecond) [Arxiv1, Arxiv2]
In the well-known exploration vs. exploitation paradigm, this strategy employs diminishing exploration alongside growing exploitation, unlike the technique of learning rate decay.

Periodic perturbation is crucial for learning the gradient in real time, but it prevents us from reaching the exact optimum.

Our algorithm allows the perturbation amplitude to decay slowly while increasing the learning rate. This results in zero-bias optimization at a desired rate of convergence in an ideal noise-free scenario.

Experimental setup for light-seeking by a unicycle robot

Aim is to navigate the vehicle to the point of maximum light intensity using only real-time light sensor data (no camera, no position data, no data storage) [Arxiv]
The forward velocity decreases over time, while the angular velocity is updated based on sensor data.
The vehicle explores the region less (with fading oscillation) and exploits the gradient estimate more (with a growing learning rate) as it approaches the source.

Light seeking with classical optimizer

Light seeking with unbiased optimizer (unbiased in theory, slight bias in practice)

Experimental setup for maximum power point tracking in a solar energy system

There exists a relationship between the power and voltage of a solar panel.
The Power-Voltage curve varies with the irradiance and temperature.
Aim is to maximize the power production by seeking the optimal voltage.
The ongoing work involves implementing the unbiased model-free algorithm to accurately determine this optimal voltage value.

Accelerated and Model-free Control for Surgical Continuum Robots

Aim is to control a continuum robot during a surgical operation.
Challanges are the uncertainty of Jacobian and unknown physiological disturbances such as respiration and heartbeat.
Developed two control strategies: one with adaptive disturbance estimator for close tracking of periodically oscillating target [Automatica], and another with time-varying gains and perfect tracking of arbitrary disturbances [ACC23]

A concentric tube robot consisting of three nested precurved tubes, subjected to periodic disturbances

A neural network structure for the estimation of the Jacobian matrix

A numerical simulation of perfect tracking of a periodically oscillating target

Control of String Actuated Systems

Aim is to stabilize the payload despite unknown disturbances and uncertainty in the cable model [Automatica]
The sensors are located only at the top boundary (where the actuator operates).
The wind disturbance (with unknown frequency, phase, and amplitude) is estimated perfectly and rejected.

Illustration of a crane system, where the payload is harmonically perturbed due to wind effects, and the aim is to stabilize the payload

Controller is OFF

The cable oscillates due to persistend wind disturbance.

Controller is ON

The controller stabilizes the payload by compensating for disturbances.

Disturbance Estimation

The unknown periodic disturbance is estimated and fed into the controller loop.

Other Projects

Beamforming

Multi-agent coordination for distributed beamforming to transmit information in the intended direction [ICASSP20, ACC20]

Time-Delay Systems

Stabilization and optimization with mobile agents in the presence of arbitrarily long communication delays [Automatica, TAC]

Suspension Systems

Quarter-car model with an active suspension system to maintain zero vertical acceleration of the vehicle body for passenger comfort, despite an unknown sinusoidal road profile and partial measurements [ACC19]

Diffusive Systems

Accelerated optimization of performance/cost functions with partial differential equation (PDE) systems, including batteries, bioreactors, drilling systems [TAC]

Pool Boiling Systems

As heat flux increases during boiling, bubbles form and rise. Beyond a critical heat flux, a vapor film insulates the heater, causing it to overheat and burn out. The goal is to seek this critical heat flux for effective cooling of electronic chips [extension of Arxiv is underway]