Deep Q-Managed: a new framework fror Multi-Objective Deep Reinforcement Learning

Abstract

The Deep Q-Managed algorithm, proposed in this work, represents a significant advancement in the field of multi-objective reinforcement learning (MORL). This new strategy employs an updated technique for hybrid multi-objective optimization, which offers a mathematical guarantee that all policies belonging to the Pareto Front can be found, excelling in the acquisition of non-dominated multi-objective policies within environments characterized by deterministic transition functions. Its flexibility extends to scenarios where the Pareto Front exhibits convex, concave, or mixed geometric complexities, making it a versatile solution for a wide array of real-world applications. Our proposal is validated using the traditional MORL benchmarks with different configurations of the Pareto front. The quality of the policies found by our algorithm was compared with prominent approaches in the literature. The outcomes of the proposed strategy establish the Deep Q-Managed algorithm as a worthy contender for tackling challenging, multi-objective problems.