A team of researchers from the University of Warsaw recently developed an AI robot to play Hearthstone from Blizzard. This is not the first machine designed to play this game, but it could very well serve as a reference for everything that follows.
Hearthstone is a trading card game, such as Magic the Gathering or Yu-Gi-Oh. Its rules and format make it an excellent candidate for exploration by machine tool developers. Like poker, players have to work with shared and limited information. The player who outperforms the other player tends to win despite the fact that luck also intervenes.
According to the researchers, they created a robot with the potential to be competitive in Hearthstone matches, similar to the way AlphaGo DeepMind was developed to defeat humans. The conclusion of their white paper published yesterday states:
The bot is able to win, with impressive consistency, 100% of games against the random player. He is also able to win games against players of rank Legend, which can be considered very promising. Human players reported that in many situations they felt that the bot played very well.
Online trading card games usually have an integrated "AI" in the form of a scripted bot. But the challenge of creating an unscripted robot capable of winning against highly skilled human players (without cheating) is not at all the same. A scripted bot only works in a protected environment, such as a solo mode or a tutorial.
To operate in an unstructured environment, robots are built using decision-tree algorithms to make choices on behalf of the AI or they are made with deep learning networks. The bot of the Warsaw team has both. It essentially combines a popular heuristic search algorithm called Monte Carlo Tree Search with a supervised deep formation network . The result is a bot that should be competitive between other robots and humans.
Despite its popularity among AI researchers, there is still no recognized benchmark for machine learning models designed to play Hearthstone. According to the Warsaw team:
We can conclude that the resulting agent will probably be among the most powerful Hearthstone robots currently. Although Hearthstone has become a testbed for AI, no universal benchmarking method has yet been proposed. It is therefore difficult to assess the force otherwise than by human observation, the self-game between different versions of the agent or a random reader. ]
This week, the team submitted its bot to the Hearthstone AI 2018 competition at the conference on computer intelligence and games .