Learning a Resource Scale for Collectible Card Games

Drafting in Collectible Card Games via Reinforcement Learning

“…To reduce the amount of episodes required to train our approaches, it may be useful to conduct some form of dimensionality reduction technique on the card feature vectors such as principal component analysis [Jolliffe and Cadima, 2016], auto-encoder networks [Kramer, 1991] or node embedding (such as the one used by Zuin and Veloso [2019] for Magic: the Gathering cards). An effort on input simplification was to sort the card choices and previously picked cards in a state by an arbitrary criterion, to virtually reduce the space state, as all states containing permutations of the same cards would result in the same input passed to the network.…”

“…Work in this direction can reveal what combination of attributes are more influential in the cards' power. On the inverse direction, Zuin and Veloso [2019] use machine learning to predict the mana cost of a card given its attributes. While this also aids CCG designers in balancing the game, the methodology of feature extraction for Magic: the Gathering cards introduced by the authors may also be useful in further work on deck building and game playing for the game.…”

“…Therefore, a more complex feature extraction process than the one we use must be employed in this case. We suggest an approach such as the one proposed by Zuin and Veloso [2019], which uses LSTM layers in the network to learn a compact representation for Magic: the Gathering cards from their raw text description.…”

“…In commercial CCGs, which feature tens of different abilities, this would be especially true. Dimensionality reduction techniques such as principal component analysis [Jolliffe and Cadima, 2016], auto-encoder networks [Kramer, 1991] or node embeddings [Zuin and Veloso, 2019] may find robust low-dimensional representation for cards, which could make training easier. Our initial efforts found good four-and six-dimensional card representations for LOCM.…”