multiobjective combinatorial optimization, learning-based method, retention model, deep reinforcement learning, Electronic computers. Computer science, QA75.5-76.95, Systems engineering, TA168

Multiobjective combinatorial optimization (MOCO) problems have a wide range of applications in the real world. Recently, learning-based methods have achieved good results in solving MOCO problems. However, most of these methods use attention mechanisms and their variants, which have room for further improvement in the speed of solving MOCO problems. In this paper, following the idea of decomposition strategy and neural combinatorial optimization, a novel fast-solving model for MOCO based on retention is proposed. A brand new calculation of retention is proposed, causal masking and exponential decay are deprecated in retention, so that our model could better solve MOCO problems. During model training, a parallel computation of retention is applied, allowing for fast parallel training. When using the model to solve MOCO problems, a recurrent computation of retention is applied, enabling quicker problem-solving. In order to make our model more practical and flexible, a preference-based retention decoder is proposed, which allows generating approximate Pareto solutions for any trade-off preferences directly. An industry-standard deep reinforcement learning algorithm is used to train RM-MOCO. Experimental results show that, while ensuring the quality of problem solving,the proposed method significantly outperforms some other methods in terms of the speed of solving MOCO problems.