Agent Based Modeling (ABM) and AI integration for smart tourism simulations

The ability to predict visitor demand at popular points of interest (POIs) and to understand tourists' visiting patterns in general is of vital importance for tourism management. We present an approach that integrates two complementary methods - agent based modeling (ABM) simulations and machine learning (ML) to enable accurate and realistic simulations of tourist movement and visiting of POIs. The ML model that predicts the next destination in the tourists' visiting sequence was trained on POI check-in data, that records tourist entrances into different attractions, using the XGBoost method.We compare different feature engineering set-ups and propose an approach for encoding the visiting history of each tourist so that it could be used in the prediction process. The model was trained and validated on 2017 data for Salzburg Card users and tested for the years 2018-2021. The results show that a large training set can yield short-term predictions with up to 75% accuracy. However, the later years are constantly predicted with lower accuracy (44%) regardless of the training set size.We also showcase the ability of our approach to produce realistic simulations of tourist visiting patterns by simulating 20 consecutive days of tourist visits in the city of Salzburg. Compared to the baseline method that makes tourists choose POIs based on popularity, and the random choice of the POIs, our ML prediction model was the only one that managed to learn different visiting patterns for different days of the week. It was also the only method that successfully learned the logical constraints of ride-type POIs where tourists usually have to take the upward ride first before coming down.