Banca de DEFESA: LUIZ PAULO DE SOUZA MEDEIROS

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : LUIZ PAULO DE SOUZA MEDEIROS
DATE: 11/09/2026
TIME: 09:30
LOCAL: Sala Virtual
TITLE:

A Multi-Agent Reinforcement Learning-Based Architecture for Multi-Objective Optimization

KEY WORDS:

Multi-objective reinforcement learning; multi-agent systems; multi-objective optimization; Pareto front; compromise policies.

PAGES: 130
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

Multi-objective optimization based on reinforcement learning aims to learn policies capable of handling conflicting objectives in sequential decision-making problems. Despite significant advances in the field, important challenges remain regarding the learning of multiple compromise policies without prior knowledge of the decision-maker's preferences and the high computational cost of environment exploration. This thesis proposes an Architecture for Multi-Objective Optimization Based on Multi-Agent Systems and Reinforcement Learning, whose main characteristic is the organization of the learning process into two complementary and cooperative stages. During the first stage, explorer agents cooperate to discover terminal states and build a shared representation of the environment through knowledge synchronization and aggregation mechanisms. During the second stage, exploiter agents reuse this shared knowledge to simultaneously learn multiple compromise policies associated with different preference vectors. This organization decouples environment discovery from Pareto frontier approximation, reducing redundant learning while promoting knowledge reuse. The proposed architecture was implemented using Q-learning and evaluated in the Deep Sea Treasure benchmark through an extensive experimental study investigating the influence of the main exploration and exploitation hyperparameters on success rate, convergence time, and generated knowledge. The results demonstrate that cooperative exploration improves the discovery of terminal states and that separating exploration from exploitation enables the learning of multiple compromise policies from a single exploratory phase. The main contribution of this research is a cooperative architecture for multi-objective optimization that employs parallelism not merely to accelerate learning, but as a mechanism for shared knowledge construction and the simultaneous generation of multiple compromise policies

COMMITTEE MEMBERS:
Presidente - 347628 - ADRIAO DUARTE DORIA NETO
Externo à Instituição - FRANCISCO CHAGAS DE LIMA JUNIOR - UERN
Externo à Instituição - IGNACIO SANCHEZ GENDRIZ - UFRN
Externo à Instituição - THIAGO HENRIQUE FREIRE DE OLIVEIRA - IFRN
Externo à Instituição - VINICIUS PONTE MACHADO - UFPI