Abstract
Conversational recommender systems support human-computer interaction strategies in order to assist online tourists in the important activity of dynamic packaging, i.e., in building personalized travel plans and in booking their holidays. In a previous paper, we presented a novel recommendation methodology based on Reinforcement Learning, which allows conversational systems to autonomously improve a rigid (non-adaptive) strategy in order to learn an optimal (adaptive) one. We applied our approach within an online travel recommender system, which is supported by the Austrian Tourism portal (Austria.info). In this paper, we present the results of this online evaluation. We show that the optimal strategy adapts its actions to the served users, and deviates from a rigid default strategy. More importantly, we show that the optimal strategy is able to assist online tourists in acquiring their goals more efficiently than the rigid strategy, and is able to increase the willingness of the users in accepting several of the system’s offers.