A novel framework for a stand-alone Orienteering Problem reinforcement learning-driven beam search
S Selvakumaran, Kumarasamy Saravanan, S Sivankalai, Subodh Kumar Suman, Vimala G, P Vithiya · 2025
This study improves the orienteering algorithm since neural network-based routing strategies have crucial issues. Most conventional approaches use node coordinates, but they use encoder-decoder patterns that don't understand the self-referential nature of routing tasks, resulting in low accuracy during the first steps of node selection and making them impractical. A hybrid technique using a variant beam search algorithm and a learnt heuristic function with higher quality and shorter computation time is offered to overcome these difficulties. Our solution uses an attention-based functional neural network to learn structural dependence between nodes via distance matrices. The reinforcement learning infrastructure trains the heuristic function to enhance node choices and sequences. Experiments show that the single- and multi-label temporal correlation models surpass existing benchmarks on numerous benchmark datasets. In particular, the suggested model generates near optimum solutions with 98.7% accuracy and beats the state-of-the-art technique by 4.3%. Using the suggested framework, convergence is 30% quicker and computing efficiency increases. These findings show that the strategy is beneficial and more efficient, accurate, and practical than current approaches for handling routing difficulties.