Article
Computer Science, Artificial Intelligence
Lei Chen, Guixiang Zhu, Weichao Liang, Youquan Wang
Summary: Trip recommendation is an intelligent service that offers personalized itinerary plans to tourists in unfamiliar cities, considering temporal and spatial constraints. In this article, we propose MORL-Trip, a Multi-Objective Reinforcement Learning approach, to address the challenges of capturing users' dynamic preferences and enhancing the diversity and popularity of personalized trips. MORL-Trip models the recommendation as a Markov Decision Process and incorporates sequential, geographic, and order information to learn user's context. It also introduces a composite reward function to reinforce accuracy, popularity, and diversity as principal objectives.
EXPERT SYSTEMS WITH APPLICATIONS
(2023)
Article
Computer Science, Information Systems
Fei Zou, Gary G. Yen, Lixin Tang, Chunfeng Wang
Summary: Dynamic Multi-objective Optimization Problem (DMOP) is a major real-world optimization problem, and efficiently tracking the movement of Pareto front over time is crucial. This paper introduces RL-DMOEA, a reinforcement learning-based dynamic multi-objective evolutionary algorithm, which effectively improves convergence and diversity of the algorithm by adapting to different severity of environmental changes through three response mechanisms.
INFORMATION SCIENCES
(2021)
Article
Engineering, Electrical & Electronic
Amir Ramezani Dooraki, Deok-Jin Lee
Summary: This paper introduces a self-trained controller for autonomous navigation in challenging static and dynamic environments. The controller utilizes deep reinforcement learning and multiple rewards to generate continuous actions while avoiding obstacles.
Article
Multidisciplinary Sciences
Menghao Wu, Yanbin Gao, Pengfei Wang, Fan Zhang, Zhejun Liu
Summary: This paper proposes a multi-dimensional action control approach based on reinforcement learning, suitable for tasks with continuous action spaces. It adopts a hierarchical structure with high and low-level modules, where low-level policies output concrete actions based on training data.
Article
Computer Science, Information Systems
Fangyuan Zhao, Xuebin Ren, Shusen Yang, Peng Zhao, Rui Zhang, Xinxin Xu
Summary: Multi-objective reinforcement learning (MORL) has potential in solving complex decision problems. We propose a novel probabilistic algorithm PMORL for efficient policy optimization and a federated MORL algorithm Fed-PMORL with client-level differential privacy for privacy protection in distributed settings.
INFORMATION SCIENCES
(2023)
Article
Computer Science, Artificial Intelligence
Manel Rodriguez-Soto, Maite Lopez-Sanchez, Juan A. Rodriguez-Aguilar
Summary: This paper addresses the problem of value alignment in multi-agent systems by proposing an approach to establish an ethical environment where agents learn to behave ethically while pursuing their individual objectives. The authors contribute by introducing a framework of Multi-Objective Multi-Agent Reinforcement Learning and formalizing a family of ethical Multi-Objective Markov Games (MOMGs) that guarantee ethical behavior learning. They also specify a process to build single-objective ethical environments, simplifying learning in multi-agent systems. As an illustration, an ethical variation of the Gathering Game is presented, demonstrating how agents learn to compensate social inequalities by aligning with the moral value of beneficence.
NEURAL COMPUTING & APPLICATIONS
(2023)
Article
Computer Science, Information Systems
Lior Hirsch, Gilad Katz
Summary: NEON is a method for network pruning using deep reinforcement learning. It achieves higher generality and reduces overfitting by training on a large set of architectures, and it becomes more efficient through offline training and a novel reward function.
INFORMATION SCIENCES
(2022)
Article
Computer Science, Artificial Intelligence
Amirhossein Asgharnia, Howard Schwartz, Mohamed Atia
Summary: This paper proposes a fuzzy multi-objective reinforcement learning algorithm that can solve multi-objective problems and handle continuous state-action domains.
Article
Computer Science, Theory & Methods
Fuhong Song, Huanlai Xing, Xinhan Wang, Shouxi Luo, Penglin Dai, Ke Li
Summary: This paper addresses the challenge of offloading dependent tasks in multi-access edge computing by proposing a new multi-objective reinforcement learning algorithm. By optimizing multiple conflicting objectives of user utility, it achieves a balance between application completion time, device energy consumption, and edge computing fees.
FUTURE GENERATION COMPUTER SYSTEMS-THE INTERNATIONAL JOURNAL OF ESCIENCE
(2022)
Article
Automation & Control Systems
SenPeng Chen, Jia Wu, XiYuan Liu
Summary: The paper proposes a novel method EMORL for hyperparameter optimization based on multi-objective reinforcement learning, which successfully addresses some limitations in traditional hyperparameter optimization. By combining accuracy and latency as a multi-objective reward, the policy update is effectively guided, resulting in improved optimization efficiency.
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE
(2021)
Article
Computer Science, Artificial Intelligence
Yupeng Han, Hu Peng, Changrong Mei, Lianglin Cao, Changshou Deng, Hui Wang, Zhijian Wu
Summary: This paper proposes a new multistrategy multiobjective differential evolutionary algorithm, RLMMDE, to solve the exploration and exploitation dilemma in multiobjective optimization problems (MOPs). The algorithm utilizes a multistrategy and multicrossover DE optimizer, an adaptive reference point activation mechanism based on RL, and a reference point adaptation method. Experimental results show that RLMMDE outperforms some advanced MOEAs on benchmark test suites and practical mixed-variable optimization problems.
KNOWLEDGE-BASED SYSTEMS
(2023)
Article
Computer Science, Artificial Intelligence
Luona Wei, Yuning Chen, Ming Chen, Yingwu Chen
Summary: The paper proposes a deep reinforcement learning and parameter transfer based approach (RLPT) for tackling the multi-objective version of the agile earth observation satellite scheduling problem (MO-AEOSSP). RLPT decomposes the problem into scalarized sub-problems and applies neural networks and reinforcement learning to generate high-quality schedules. Experimental results show that RLPT outperforms traditional algorithms in terms of solution quality, distribution, and computational efficiency.
APPLIED SOFT COMPUTING
(2021)
Article
Robotics
Lingfeng Sun, Haichao Zhang, Wei Xu, Masayoshi Tomizuka
Summary: In this work, the potential of improving multi-task training and transferring in the reinforcement learning setting is investigated. Challenges are identified and a parameter-compositional formulation is proposed for transferring. Ways to improve the training of multi-task reinforcement learning are explored as the foundation for transferring. Several transferring experiments on various manipulation tasks are conducted, demonstrating improved performance in the multi-task training stage and effective transferring in terms of both sample efficiency and performance.
IEEE ROBOTICS AND AUTOMATION LETTERS
(2023)
Article
Computer Science, Artificial Intelligence
Lilu Zhu, Feng Wu, Yanfeng Hu, Kai Huang, Xinmei Tian
Summary: In this paper, we propose a heuristics multi-objective task scheduling framework called AC-CCTS based on reinforcement learning. The framework solves the problems of single objective and local convergence in traditional task scheduling methods and reduces the cost of experiential learning. Experimental results show that AC-CCTS outperforms other meta-heuristic scheduling methods and reinforcement learning algorithms in terms of resource utilization efficiency and convergence stability in container-based cloud task scheduling.
NEURAL COMPUTING & APPLICATIONS
(2023)
Article
Engineering, Industrial
Jinling Leng, Xingyuan Wang, Shiping Wu, Chun Jin, Meng Tang, Rui Liu, Alexander Vogl, Huiyu Liu
Summary: This study investigates a multi-objective resequencing scheduling problem in automotive manufacturing systems. By resequencing cars based on color orientation, costs of color changes and operational costs for paint shops can be reduced. The proposed multi-objective deep-Q-network algorithm determines the Pareto frontier and improves training efficiency through reward shaping and the design of a 2D-folded-normal distribution for preference sampling. Experimental results show that the proposed approach outperforms other algorithms in terms of time, performance, neural network convergence, and the diversity of the Pareto frontier.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
(2023)