Fuzzy Multiset Clustering for Metagame Analysis

Developing agents for automated game playing is a demanding task in the general game production cycle. Especially the involvement of frequent balance changes after the release, which for example often occur in collectible card games, require constant updates of the developed agent. The game’s developers need to continuously analyze and understand the current meta-game for adjusting the agent’s parameters, making balance changes to the game, and, thereby, sustaining the satisfaction of its player base. The underlying analysis largely depends on evaluating players’ play traces. Necessary adjustments to the agent’s and the game’s parameters are taken care of by the game’s developers. This paper proposes a first step in automatically observing the current state of a collectible card game, which will assist the developers in their understanding of established deck archetypes and, therefore, speed up the update cycle. Fuzzy multisets are used for modeling decks and frequently occurring subsets of cards. We propose the definition of a (fuzzy) multiset centroid to uniquely represent the cluster and its contained decks and show that it is better able to match the deck archetype than the often reported deck core. The proposed clustering procedure identifies deck archetypes and keeps track of its common variants in the current meta-game. We evaluate the approach by comparing the result of our clustering procedure with a hand-labeled data set and show that it is able to reproduce clusters of similar quality to a labeling provided by experts.