by Grail TeamHelping AI to Play Hearthstone
AAIA’17 Data Mining Challenge: Helping AI to Play Hearthstone (https://knowledgepit.fedcsis.org/contest/view.php?id=120) took place between March 23 and May 15, 2017. It was organized under the auspices of the 12th International Symposium on Advances in Artificial Intelligence and Applications (AAIA’17, https://fedcsis.org/2017/aaia) and sponsored by Silver Bullet Solutions (http://www.silverbullet.pl/en/home/) and Polish Information Processing Society (http://www.pti.org.pl/English-Version).
This initiative was a part of our Grail project. The main objective in the competition was to construct a prediction model which would be able to foresee who is going to win, using only information about a single game state. Such models can be utilized in practice to improve performance of game state search heuristics, such as Monte Carlo Tree Search (MCTS), e.g. by reducing simulation times – instead of simulating until the end of the game we may simulate only k steps ahead and estimate the result of the game using the prediction model.
The dataset provided to participants contained examples of game states extracted from simulated Hearthstone play outs between weak AI players (i.e. the agents which were used to generate the data were choosing their in-game decisions at random). The participants were asked to predict winning chances of the first player from game states belonging to the test set and submit their predictions to the Knowledge Pit competition platform. In order to give participants a freedom of choosing a representation of the data which they want to use, the datasets were provided in two formats: in a tabular format (with a simplified representation) and as raw JSON files (with detailed game states).
Even though AAIA’17 Data Mining Challenge lasted for less than two months, it attracted attention of many researchers from domains of machine learning and artificial intelligence in video games. By the end of competition there were 296 teams from 28 countries registered in the challenge. Among them, 188 teams submitted at least one solution to the public leaderboard and 114 teams described their solution in a report uploaded to the Knowledge Pit platform. In total, we received 4067 submission, which makes this competition the most popular one among challenges organized at Knowledge Pit to this day. The official results were announced during the award ceremony at conference banquet. Two top ranked teams received money prizes. Moreover, seven teams were invited to present their solutions during a special session at the conference.
During the conference we prepared a stand where participants were able to test their Hearthstone skills by playing against our AI bot. The games were played using our simple web app UI (http://126.96.36.199/grailstone). We received a lot of positive feedback regarding performance of the bot. From a few dozens of games played during the conference, a human player won only once, with an experienced Hearthstone player. Technically, our bot was based on the MCTS algorithm supported by a value network (i.e. a prediction model for the assessment of game states) and a simple policy network (a prediction model for the evaluation of possible actions). The first model was trained on data similar to the sets used in the competition. To train the second model, we used data about actions chosen by different artificial bots. This approach was a consequence of our limited access to data about games played between experienced human opponents. With an access to such data our AI bot would become even more challenging.
Download article here.