Seminar M.Sc Students
Dynamic Budget Allocation for an Advertising Campaign
Jilnar Manassah, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Evgeni Khmelnitsky
Abstract: Digital advertising campaigns require advertisers to distribute a limited budget over time while facing uncertain, delayed and nonlinear returns. Spending decisions are interdependent: allocating more budget early in a campaign may improve immediate exposure but reduce the resources available for later periods. Moreover, the effect of advertising expenditure may persist over time and typically exhibits diminishing marginal returns. This research studies the dynamic budget-allocation problem of determining how much to spend in each period to maximize the expected net profit of an advertising campaign, subject to total-budget and per-period spending constraints.
We formulate the dynamic budget-allocation problem in both discrete and continuous time using a nonlinear revenue-response model in which current and past advertising expenditures contribute to revenue through delayed, Poisson-shaped response weights. The discrete time formulation represents revenue as the accumulated effect of expenditures across periods, while the equivalent continuous time formulation describes revenue evolution through a system of differential equations. For a particular parameterization, this system resembles a controlled physical system in which advertising expenditure acts as an external input affecting revenue dynamics. We also consider a more general continuous time model that provides greater flexibility for estimating the revenue dynamics and the nonlinear effect of expenditure from data.
The proposed framework is evaluated using real-world advertising campaign data provided by Playtika. The empirical analysis examines how accurately the assumed response model predicts campaign revenue and compares the optimized allocation with the observed spending policy. While the Poisson-based model provides a strong fit for some campaigns, its performance varies across the dataset. To address this limitation, we extend the framework to an online setting in which model parameters are repeatedly updated as new campaign observations become available. We further propose a bi-level regression framework that incorporates additional campaign features influenced by advertising expenditure, with the goal of improving revenue prediction while preserving expenditure as the primary decision variable.
Bio: Jilnar Manassah is a M.Sc. student in Industrial Engineering at Tel Aviv University, specializing in Data Science and Artificial Intelligence. Her research focuses on the dynamic allocation of advertising budgets using optimization, statistical learning and data-driven methods.
Multiagent LLM Debate Violates Classical Opinion Dynamics
Avigail Kollmann, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Dr. Ilai Bistritz
Abstract: Multi-agent debate (MAD) systems have emerged as a promising approach for improving the reasoning capabilities of large language models (LLMs), yet the impact of network topology on their performance remains understudied. We systematically investigate how different communication topologies affect both accuracy and token efficiency in multi-agent LLM systems across diverse reasoning tasks. We evaluate five distinct topologies using four language models on four datasets spanning mathematical reasoning, grade-school problems, conversational AI, and reading comprehension. Additionally, we introduce a simplified probabilistic model that captures agent opinion dynamics through majority-based updates with noise parameters based on p and q probabilities. We show that by choosing appropriate p and q values, we can simulate the LLM model behavior for chain-of-thought models. Nevertheless, for MAD, no p and q can approximate the LLM behavior. In fact, for every different task or LLM model, a different topology achieves the best accuracy or the least tokens spent. Our results highlight that LLM agents do not align with the classical theory of consensus over graphs.
Bio: Avigail Kollmann is a M.Sc. student in Industrial Engineering at Tel Aviv University, specializing in Data Science and Artificial Intelligence. Her research focuses on the dynamics of consensus and reasoning in multi-agent LLMs using systematic topological evaluation and probabilistic modeling.
“Balancing effectiveness and equity in facility location modeling under uncertainty”
Hadar Engel, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Dr. Reut Noham
Abstract: Decisions about where to locate public facilities, such as healthcare centers or humanitarian aid hubs, are a central determinant of service accessibility and system performance, particularly under demand uncertainty and limited resources. In this work, we examine how equity can be systematically incorporated into stochastic facility location decisions, moving beyond traditional models that prioritize efficiency or expected effectiveness alone. It introduces the Equity-Driven Stochastic Facility Location (ED-SFL) framework, which explicitly balances effectiveness, efficiency, and equity both within and across demand scenarios. Building on social-welfare measures developed for deterministic settings, we adopt a recently introduced scenario-balanced objective function and apply it, for the first time, to stochastic facility location and mixed integer linear program. Through rigorous mathematical modeling and the development of exact and heuristic solution methods, our work aims to provide theoretical insights and practical tools for designing equitable and more robust public-sector service systems.
