A Profile-Based ‘GrEvolutionary’ Hearthstone Agent

In the last few years, the Hearthstone AI international Competition has been gaining fame among the scientific community. Several different entries have been presented using varied approaches. One of the best, EVA, was based on a Greedy approach combined with an Evolutionary Algorithm. However, almost all the proposals were designed to work in a general way, i.e. for any of the possible heroes. This generalisation presents a flaw, since the exclusive cards per hero are not really exploited, nor their potential different behaviour profiles. This paper follows a similar philosophy to EVA, also hybridizing Greedy + Evolutionary algorithms, but having in mind three different, and extended among the community, archetypes or profiles: Aggro, Control and Midrange. Thus, three different behaviours have been optimized aiming to create a more specialized agent able to use an Artificial Intelligence engine depending on the hero to play with. To prove the value of the approach several experiments have been conducted, comparing the evolved agents with EVA in many different matches using three different heroes. The results show an improvement over EVA for the three profile-based agents, as well as an excellent performance when combined with a MonteCarlo Tree Search approach.