Scholarlink

<p>The key to solving constrained multi-objective optimization problems (CMOPs) is to maintain an effective balance among diversity, convergence, and feasibility. The integration of evolutionary algorithms (EAs) with Deep Q-Networks (DQN) provides a promising framework for addressing CMOPs. Accordingly, this paper proposes a constrained multi-objective optimization algorithm based on DQN and time series prediction (CMODQN-TSP). In CMODQN-TSP, the feasibility, diversity, and convergence of the population are defined as the state, while candidate operators are treated as actions. The improvement of the population state is used as the reward signal. The DQN estimates the Q-values for a given state to guide operator selection and promote population evolution toward the constrained Pareto front. In addition, a time series prediction mechanism is incorporated, where an Autoregressive eXogenous (ARX) model and a Gaussian Mixture Model-based local search are used to predict the next generation of solutions within the feasible domain according to population distribution characteristics. By integrating deep reinforcement learning with predictive modeling, the proposed algorithm improves adaptability and search performance for complex CMOPs. To evaluate the effectiveness of CMODQN-TSP, five state-of-the-art constrained multi-objective evolutionary algorithms (CMOEAs) are selected for comparison on four benchmark test suites and a real-world problem. The experimental results demonstrate that CMODQN-TSP achieves superior performance compared with the other algorithms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Constrained multi-objective optimization based on deep Q-Network and time series prediction

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Peng Zhao,
  • Jie Cao,
  • Jianglin Zhang,
  • Zuohan Chen,
  • Chengzhi Liu

摘要

The key to solving constrained multi-objective optimization problems (CMOPs) is to maintain an effective balance among diversity, convergence, and feasibility. The integration of evolutionary algorithms (EAs) with Deep Q-Networks (DQN) provides a promising framework for addressing CMOPs. Accordingly, this paper proposes a constrained multi-objective optimization algorithm based on DQN and time series prediction (CMODQN-TSP). In CMODQN-TSP, the feasibility, diversity, and convergence of the population are defined as the state, while candidate operators are treated as actions. The improvement of the population state is used as the reward signal. The DQN estimates the Q-values for a given state to guide operator selection and promote population evolution toward the constrained Pareto front. In addition, a time series prediction mechanism is incorporated, where an Autoregressive eXogenous (ARX) model and a Gaussian Mixture Model-based local search are used to predict the next generation of solutions within the feasible domain according to population distribution characteristics. By integrating deep reinforcement learning with predictive modeling, the proposed algorithm improves adaptability and search performance for complex CMOPs. To evaluate the effectiveness of CMODQN-TSP, five state-of-the-art constrained multi-objective evolutionary algorithms (CMOEAs) are selected for comparison on four benchmark test suites and a real-world problem. The experimental results demonstrate that CMODQN-TSP achieves superior performance compared with the other algorithms.