Optimization of digital advertising campaigns in non-stationary environments through a reinforcement learning algorithm

As consequence of the spread of Internet, digital advertising is becoming one of the most rewarding marketing instruments, e.g. providing about 97% of the Google's revenue for an amount of about $60 billions. Initially, digital campaigns were managed manually by marketing experts, while recently, to face their increasing complexity, campaigns require an automatic optimisation. In this work, we developed for the first time an optimisation algorithm for event marketing, which is an unexplored scenario in the scientific literature for the non-stationary of its environment. Using reinforcement learning techniques based on states and actions, we propose an online algorithm based on users' populations which sets the daily budget for multiple advertisement channels, e.g., search engines and social networks, and multiple campaigns per channel in order to maximise the number of tickets sold for events. We experimentally evaluate the algorithm to advertise the Salon du Chocolat - Milan 2017. As support of the algorithm an expert in the field positively reviewed the results.