Work

Deep Q-Managed: a new framework for multi-objective deep reinforcement learning

Year: 2025

Type: article

Abstract: This paper introduces Deep Q-Managed, a novel multi-objective reinforcement leaning (MORL) algorithm designed to discover all policies within the Pareto Front. This approach enhances multi-objective optimization by integrating deep leaning techniques, including Double and Dueling Networks, to effectively mitigate the curve of dimensionality and overestimation bias. Deep Q-Managed demonstrates high proficiency in attaining non-dominated multi-objective policies across deterministic episodic environments, adapting to convex, concave, or mixed Pareto Front complexities. Experiments on traditional MORL benchmarks (Deep Sea Treasure, Bountiful Sea Treasure, and Modified Bountiful Sea Treasure) show it consistently achieves maximum hypervolume values (e.g., 1,155 for DST, 3,352 for BST, and 2,632 for MBST) and locates all Pareto Front points. While robust and versatile for practical applications in robotics, finance, and healthcare, this study's validation is currently confined to deterministic episodic settings, with stochastic environments reserved for future work.

Source: Frontiers in Artificial Intelligence

DOI: 10.3389/frai.2025.1683323

Authors: Richardson Menezes, Thiago Henrique Freire de Oliveira, Luiz Paulo de Souza Medeiros, Adrião Duarte Dória Neto

Institutions: Universidade Federal do Rio Grande do Norte, Centro Universitário Paulistano, Instituto Federal do Rio Grande do Norte

Language: English

Topic: Reinforcement Learning in Robotics

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,353

Funder: Conselho Nacional de Desenvolvimento Científico e Tecnológico