How Can AI Design the Perfect Route to Explore a Heritage City?

Picture strolling through Warin Chamrap, a small town packed with history, without getting lost or wasting time. Researchers have built a system called DRL‑IMVO‑GAN, blending artificial intelligence and evolutionary algorithms to craft tailor‑made tourist routes. At its core is Deep Reinforcement Learning (DRL), which learns to pick the most interesting points of interest (POIs) by watching how visitors move in a digital twin of the city. Next, the Improved Multiverse Optimizer (IMVO) explores countless possible solutions, balancing travel time, CO₂ emissions, spatial coherence, and group satisfaction. Finally, a Generative Adversarial Network (GAN) fine‑tunes the routes, adding variety and realism. Tested on a virtual replica with 92 POIs, congestion heatmaps, and behavior‑based tourist clusters, the model outperformed seven other methods, scoring highest in cultural heritage (74.2), shortest travel time (21.3 minutes) and top satisfaction (17.5 out of 18). Moreover, it adapts to new attractions or constraint changes without retraining, showing great flexibility. In practice, DRL‑IMVO‑GAN could become a digital travel assistant, helping visitors enjoy more sustainable, personalized cultural experiences even in lesser‑known cities.

The DRL‑IMVO‑GAN framework rests on three core components. Deep Reinforcement Learning employs a deep neural network to learn a POI selection policy through a trial‑and‑error process simulated in the digital twin. The reward function integrates multiple objectives: heritage score (based on historical and cultural value criteria), travel time, CO₂ emissions estimated with transport energy consumption models, spatial coherence (measured via geodesic proximity metrics), and group satisfaction (inferred from interest‑diversity metrics). The Improved Multiverse Optimizer, an evolution of the Multiverse Optimizer algorithm, leverages quantum‑physics concepts such as multiverse theory to generate candidate solutions through “wormhole” and “black‑hole” operations that balance global exploration and local exploitation. The GAN, composed of a generator and a discriminator, acts as a refiner: the generator proposes minor route adjustments (e.g., swapping nearby POIs), while the discriminator assesses quality against a predefined set of optimal solutions, ensuring diversity and robustness. Validation employed multi‑objective metrics: hypervolume (0.85) and Pareto Coverage Ratio (0.95) indicate broad coverage of the non‑dominated solution space. Moreover, the model demonstrated zero‑shot and few‑shot capabilities thanks to a latent POI representation that allows new points to be added without retraining the entire network. This hybrid approach combines the sequential learning ability of DRL with the global search power of IMVO and the generative flexibility of GAN, delivering a scalable solution for sustainable tourist itinerary planning in complex urban contexts.