Área: Engenharia Informática e de Computadores

Despacho de nomeação de Júri

Edital

Título da Tese: Deep Networked Agents with Applications in Adaptive Traffic Signal Control

Local da Prova:  https://tecnico-pt.zoom.us/j/99784082282 

Data: 28/10/2026

Hora: 14h00

Abstract: We investigate a certain fundamental problem in decentralized training and decentralized execution of reinforcement learning-based multi-agent systems. These agents interact sequentially with the environment, collecting different (but related) experiences, and must take compatible actions towards reaching a common (perceived) goal. Differently from their centralized training counterparts, decentralized training agents lack global system information such as its state. In this solitary existence, agents are only free to pursue self-interested goals. Thus, we consider a scheme whereby agents asynchronously use limited peer-to-peer communication with a dual purpose; to tentatively align goal evaluations, and, to exchange parameters reflecting their accumulated knowledge. This class of multi-agent system, networked agents, finds a broad range of applications, e.g., game theory, evolutionary biology, supply chain and smart infra-structure, among others. This thesis particularly addresses a problem in intelligent transportation systems: Adaptive traffic signal control (ATSC). In this context, each agent controls an intersection by selecting red/green signals using local sensors to minimize the average travel time for the network. We contribute with two (general purpose) novel deep networked agents algorithms: the first in the cooperative setting, generates factorized representations for the objective function expressing the goal. The other, builds localized approximations for the (perceived) common objective by polling neighbors for their individual goal evaluations. We showcase these algorithms on the ATSC task.