An improved transformer model with multi-head attention and attention to attention for low-carbon multi-depot vehicle routing problem

• Published in: Annals of operations research Vol. 339; no. 1-2; pp. 517 - 536
• Main Authors: Zou, Yang, Wu, Hecheng, Yin, Yunqiang, Dhamotharan, Lalitha, Chen, Daqiang, Tiwari, Aviral Kumar
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2024; Springer Nature B.V
• Subjects: Business and Management; Carbon; Carbon content; Combinatorics; Effectiveness; Encoders-Decoders; Genetic algorithms; Industrial development; Logistics; Machine learning; Operations Research/Decision Theory; Original Research; Queries; Recurrent neural networks; Sustainable development; Theory of Computation; Transformers; Vehicle routing; Transformer model; Low-carbon multi-depot vehicle routing problem; Multi-head attention mechanism; End-to-end deep reinforcement learning; GA algorithm
• ISSN: 0254-5330; 1572-9338
• DOI: 10.1007/s10479-022-04788-z

Low-carbon logistics is an emerging and sustainable development industry in the era of a low-carbon economy. The end-to-end deep reinforcement learning (DRL) method with an encoder-decoder framework has been proven effective for solving logistics problems. However, in most cases, the recurrent neural networks (RNN) and attention mechanisms are used in encoders and decoders, which may result in the long-distance dependence problem and the neglect of the correlation between query vectors. To surround this problem, we propose an improved transformer model (TAOA) with both multi-head attention mechanism (MHA) and attention to attention mechanism (AOA), and apply it to solve the low-carbon multi-depot vehicle routing problem (MDVRP). In this model, the MHA and AOA are implemented to solve the probability of route nodes in the encoder and decoder. The MHA is used to process different parts of the input sequence, which can be calculated in parallel, and the AOA is used to deal with the deficiency problem of correlation between query results and query vectors in the MHA. The actor-critic framework based on strategy gradient is constructed to train model parameters. The 2opt operator is further used to optimize the resulting routes. Finally, extensive numerical studies are carried out to verify the effectiveness and operation efficiency of the proposed TAOA, and the results show that the proposed TAOA performs better in solving the MDVRP than the traditional transformer model (Kools), genetic algorithm (GA), and Google OR-Tools (Ortools).