A reinforcement learning approach to the orienteering problem with time windows

• Adaptable experimental design to solve the Tourist Trip Design Problem. Performance, generalization and inference times suitable for real-time applications. Sets ground incorporating real data in supervised training. Improvements Pointer Network model extendable similar combinatorial problems. include recursivity Transformer one-step lookahead mask. The Orienteering Problem with Time Windows (OPTW) is a optimization problem where goal maximize total score collected from different visited locations. application of neural network models has recently shown promising results dealing problems, like Travelling Salesman A allows learning solutions using reinforcement or learning, depending on available data. After stage, it can be generalized quickly fine-tuned further improve performance personalization. advantages are evident since, real-world applications, solution quality, personalization, execution all important factors that should taken into account. This study explores use trained OPTW problem. We propose modified architecture leverages Networks better address problems related dynamic time-dependent constraints. Among its various used (TTDP). train TTDP mind, by sampling variables change across tourists visiting particular instance-region: starting position, time, scores given each point interest. Once model-region trained, infer tourist beam search. based assessment our approach several existing benchmark instances. show generalizes visit region generally outperforms most commonly heuristic, while computing realistic times.