Evaluation of hearthstone game states with neural networks and sparse autoencoding

In this paper, an approach to evaluating game states of a collectible card game Hearthstone is described. A deep neural network is employed to predict the probability of winning associated with a given game state. Encoding the game state as an input vector is based on another neural network, an autoencoder with a sparsity-inducing loss. The autoencoder encodes minion information in a sparse-like fashion so that it can be efficiently aggregated. Additionally, the model is regularized by decorrelation of hidden layer neuron activations, a concept derived from an existing regularizing method DeCov. The approach was developed for AAIA'17 data mining competition “Helping AI to play Hearthstone” and achieved 5th place out of 188 submissions.