Multi-objective Reinforcement Learning – Concept, Approaches and Applications
Linzi Zhang, Zhiquan Qi, Yong Shi
Abstract
Real-world decision-making tasks are generally complicated and require trade-offs between multiple, even conflicting, objectives. As the advent and great development of advanced information technology, it has evolved into using reinforcement learning (RL) algorithms to tackle the multi-objective decision making (MODM) problems. In this paper, we will first identify the basic concepts and factors when modelling the MODM tasks with reinforcement learning, and then review the traditional RL, such as Sarsa, Q-Learning, Policy Gradients, Actor-Critic, Monte-Carlo learning, and modern deep RL algorithms applied in this process. Furthermore, the specific practical scenarios described in MODM problems will be summarized through analyzing some typical articles. Finally, the future trends of multi-objective reinforcement learning will be discussed.