Predicting Human Card Selection in Magic: The Gathering with Contextual Preference Ranking

Supervised and Reinforcement Learning from Observations in Reconnaissance Blind Chess

“…The parameters of F are trained such that distances between these embeddings model the strength of the relationship of the inputs, with the goal that d p < d n , as illustrated in the centre of Figure 1. 2 This is commonly achieved by minimising the triplet loss.…”

“…One can also view each triplet as a contextual preference, denoted as (p t ≻ n t,i | O t ). In this way, our approach becomes a version of contextual preference ranking (CPR), which uses Siamese neural networks for preference-based decisionmaking [2]. In particular, Bertram et al used CPR to measure the synergy of cards in a collectable card game, where the Siamese network modelled how well a candidate card fits a set of previously chosen cards.…”

Player Identification and Next-Move Prediction for Collectible Card Games with Imperfect Information

“…• We introduce a dedicated dataset 1 , specifically created for studying RQ1-RQ4, which captures the entire game state of a virtual CCG at each turn in the game and links that state to a unique player and their action, resulting in nearly two million tuples of gameplay data. This is notable as the limited amount of CCG play data has previously hindered AI and gaming research (Bertram, Fürnkranz, and Müller 2021). Although there are opensource CCGs for research, such as LoCM, to the best of our knowledge, no other comprehensive datasets of gameplay information exist.…”

“…Collectible card games, also known as trading card games, typically focus on two aspects: deckbuilding and player-vs-player combat (Turkay, Adinolf, and Tirthali 2012). The specific mechanics vary between games, but in most CCGs players build their own card decks, either by selecting one card at a time from a set of pre-specified options (i.e., drafting) (Vieira, Tavares, and Chaimowicz 2020) or by creating synergistic groups to perform specific multi-turn effects from the cards that the player owns (Bertram, Fürnkranz, and Müller 2021;Bjørke and 1 The datasets for this work are available upon request. Fludal 2017).…”

“…However, behavioral stylometry and other forms of play style identification remain a significant issue for CCGs (Bertram, Fürnkranz, and Müller 2021). Classifying sequences of common moves, as well as the possible play style strategies which apply them, based on data from large numbers of players is a continuing topic of research.…”