Seminar M.Sc Students
Distributed Formation Control with Biased and Missing Distance Measurements/strong>
Adi Shalit, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Dr. Ilai Bistritz
Abstract: Formation control is a fundamental problem in multi-agent systems, where autonomous agents must converge to a desired geometric configuration. In many realistic robotic scenarios, agents do not have access to global positions, direct communication, or accurate relative measurements. This work studies a fully distributed setting in which each agent observes only noisy and biased scalar distance measurements to its neighbors, resulting in a partial and distorted view of the formation. We formulate the problem as a potential game whose objective is an expected stress-based formation cost, closely related to Euclidean Distance Matrix reconstruction from corrupted distance measurements. To handle collision singularities, we introduce a smooth repulsive cost at short distances and analyze the resulting smoothed objective under measurement noise. We propose D-FORCE, a fully distributed two-point zeroth-order algorithm that estimates descent directions using only local distance measurements, without gradients, bearings, communication, or knowledge of the bias. We prove convergence toward stationary points of the expected objective, with the minimum expected squared gradient norm vanishing at rate (O(T^(-2/5) )). Simulations show that the proposed method reconstructs formations close to the desired geometry despite persistent measurement bias.
Bio: Adi Shalit is an M.Sc. student in the School of Industrial Engineering, under the supervision of Dr. Ilai Bistritz. His research focuses on multi-agent learning, distributed optimization, and formation control, with applications to robotic systems operating under noisy, biased, and partial information. He is particularly interested in decentralized algorithms that enable autonomous agents to coordinate without centralized control or communication.
A Natural Experiment in Web Privacy: Multi-Layer Auditing of Israeli Sites Around Amendment 13
Rom Ackerman, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Eran Toch
Abstract: Modern digital platforms rely on a deeply integrated, complex infrastructure of tracking scripts, real-time bidding entities, and background pixels managed by global platforms. These third-party networks capture extensive behavioral data, often operating in the background before users can even engage with visual consent mechanisms. To address these growing data extraction practices and modernize a regulatory framework that had remained largely untouched for long period, Israel introduced Amendment No. 13. This legislation aimed to bring domestic privacy protections into closer alignment with contemporary global standards, such as the EU's General Data Protection Regulation (GDPR). The amendment enforced from August 2025, created a natural experiment: did the law actually change how Israeli websites track their users, or only how they ask? To investigate this, this research introduces a multi-layer auditing framework designed to evaluate compliance by monitoring a panel of domestic and international domains that was assembled in two waves, one prior to the amendment enforcement and one post the enforcement, to track the real-time technical adaptations triggered by the law. Our primary finding highlights an illusion of compliance: after the amendment, consent banners and policy links proliferated, yet the underlying volume of tracking barely moved. This talk presents the multi-layer measurement pipeline, the results of our analysis, and how this methodology can be incorporated into a regulator's toolbox for more effective enforcement.
Bio: Rom Ackerman holds a B.Sc. in Communication system Engineering from Ben-Gurion University of the Negev. He is currently a Tech Lead at Finiti and an M.Sc. student in the Department of Industrial Engineering at Tel Aviv University. His research focuses on systems and web privacy auditing, operating at the intersection of complex data infrastructures and regulatory frameworks.
Accelerating Response-Curve Learning in Digital Advertising via Sequential Bayesian Experimental Design
Niv Danieli, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Evgeni Khmelnitsky & Dr. Yossi Luzon
Abstract: Digital advertising campaigns require advertisers to distribute a limited budget across multiple marketing channels while facing non-linear, saturating returns. To optimize these allocations, advertisers must accurately map channel-specific response curves, a process that traditionally requires a costly and typically randomized initial exploration phase. Because every exploration decision carries real financial consequences, a prolonged learning period incurs substantial opportunity costs and budget inefficiencies. This research studies the sequential budget exploration problem, focusing on how to dynamically select daily spend levels to learn underlying advertising saturation curves as rapidly as possible. We formulate the budget exploration problem as a sequential experimental design task, using a non-linear response model parameterized by the three-parameter Hill function to capture diminishing marginal returns. Rather than exploring through unguided schedules, we implement a sequential I-optimal Bayesian Experimental Design (BED) framework. The optimization rule evaluates the analytical Fisher Information Matrix to sequentially select budget levels that minimize the integrated out-of-sample prediction variance across the continuous feasible space, prioritizing information gain to shorten the campaign's learning phase. The proposed framework is evaluated using real-world mobile application user acquisition data provided by Playtika Ltd. The empirical analysis demonstrates that I-optimal BED significantly accelerates curve learning, matching the out-of-sample prediction accuracy of random sampling while using 45% to 82% fewer observations. To adapt to market dynamics where ad performance shifts abruptly, we extend the baseline model via a rolling-window estimator to track non-stationary, episodic parameter drift. We further generalize the framework to a joint multi-campaign portfolio optimization model that coordinates budget distribution across concurrent regional segments under shared daily caps, pacing bounds, and cumulative historical spend constraints.
Bio: Niv Danieli is an M.Sc. student in Industrial Engineering at Tel Aviv University, specializing in Data Science and Machine Learning. His research focuses on sequential budget exploration and advertising saturation curve estimation using active learning and statistical modeling frameworks.
