A Neural Network Approach to Hearthstone Win Rate Prediction

Abstract

This paper describes a solution to the AAIA'18 data mining challenge, which concerns prediction of win rates for decks in Hearthstone collectible card game. A neural network model assigning win rate to decks is learned based on maximisation of log probability of observed match results. A representation of deck contents is based on a second network, which performs the role of a dual-task encoder. Two tasks learned by the encoding networks are encoding decks in such a way that the full deck can be reconstructed, and encoding individual cards so that their specific properties can be decoded. Shared representation for these tasks allows the knowledge of individual cards to be taken into account.

Looking for archetypes: Applying game data mining to hearthstone decks

“…Indeed there have been two Data Mining Challenges (AAIA'17 3 and AAIA'18 4 ) using this game as a test-bed. However, the 2017 Challenge and the derived papers [18,19] were devoted to help the AI win the game, whereas the 2018 edition and related papers [20,21] had the aim to predict the win-rates for specific decks. Regarding this, [22] used decks archetypes for the prediction of battle outcomes, creating clusters, and predicting the win-rates comparing their similarity with standard archetype decks.…”

Deep Surrogate Assisted MAP-Elites for Automated Hearthstone Deckbuilding

“…As a result, many works train AI agents to play Hearthstone [16][17][18][19][20][21][22][23][24][25]. Other works predict the result of a game given a partial log of games [26,27] or predict the archetype of a deck from the first round of the game [28]. Previous work [29] demonstrates that quality diversity algorithms can generate a collection of Hearthstone agents with diverse strategies for a single deck.…”

“…After the inner-loop runs 𝑛 iterations, we return to the outer loop. We extract all elites from the surrogate archive 𝑀 𝑠 and evaluate them on the Hearthstone to obtain ground-truth predictions 𝑓 , m, and 𝛼 (lines [22][23][24][25][26][27]. These decks and their ground-truth data get added to both the training data buffer for the surrogate model (line 24) and the outer loop's ground truth archive (line 25).…”

The Many AI Challenges of Hearthstone

“…The same approach, training on human or machine gameplay logs, would also work for constructing an action recommender. Win-rate predictors for Hearthstone have been the focus of previous research [29], driven specially by the AAIA Data Mining Competition [31]. In general, a good starting point for this type of gameplay assistance systems would be a game-playing algorithm; a significant research challenge though, is which information to present to the player and how.…”