Deep Reinforcement Learning Using Graph Conv-LSTM for Solving Multi-Objective Combinatorial Optimization Problems

This paper introduces MOGCL (Multi-Objective Graph Convolutional Long Short-Term Memory with Attention), a novel model that combines dynamic graph Convolutional Long Short-Term Memory networks (Conv-LSTM) and attention mechanisms for addressing multi-objective combinatorial optimization problems. The algorithm employs an encoder-decoder architecture, leveraging Conv-LSTM's characteristics to capture more spatio-temporal features, thus enabling a comprehensive understanding of the graph's topological and latent node relationships. Using the Multi-Objective Traveling Salesman Problem (MOTSP) as a case study, the model utilizes Graph Attention Networks to capture intricate city node interactions, learning to efficiently filter and transmit information, which enhances the training speed and performance of the MOGCL model. Additionally, two novel methods were introduced to the conventional weight handling techniques, expanding the options for weight processing and improving the model's diversity generalization capabilities. Experimental results comparing the algorithm with other heuristic and learning-based methods demonstrate its efficient processing of MOTSP and significant improvements in effectiveness and model training efficiency, particularly in resource-constrained environments, leading to a substantial reduction in model training time.