Evolutionary Deckbuilding
in Hearthstone
Evolutionary Deckbuilding
in Hearthstone
Pablo García-Sánchez, Alberto Tonda,
Giovanni Squillero, Antonio Mora and JJ Merelo
Pablo García-Sánchez, Alberto Tonda,
Giovanni Squillero, Antonio Mora and JJ Merelo
The Problem
Collectible Card Games
Deckbuilding
Complex and unpredicted
reactions
Balancing the game
(Source: Cody Durkin, www.mastermarf.com, Creative
Commons Attribution-Noncommercial-Share-Alike 3.0
license)
Proposal
● Methodology to create decks for CCGs
using an Evolutionary Algorithm.
● Decks (solutions) are encoded as vectors
● Fitness (metric of quality) is measured
against human made decks
● Proof-of-concept using HearthStone
Some Concepts in CCG’s
Types of deck: Aggro, Combo, Control
Metagame
Archetypes (Dragon Priest, Secret Paladin…)
Season
Methodology
1) Select a finished season
2) Select a wide range of enemies for evaluation
3) Select the set of cards to use
4) Run several battles per evaluation
Opponent Selection
Deck Name Games Won Games Lost Win/Lose Ratio
Aggro Paladin 182 74 0.71
Mage Tempo 177 79 0.69
Shadow Madness Priest 152 104 0.59
Midrange Hunter 143 113 0.55
Mech Shaman 119 137 0.46
Oil Rogue 106 150 0.41
Control Warrior 104 152 0.40
Midrange Druid 85 171 0.33
Warlock MalyLock 83 173 0.32
Experimental Setup
50 Generations
10 individuals
Generational replacement
Using uGP framework
16 battles against each opponent
(128 per individual evaluation)
Conclusions and future work
Proposal seems promising: useful cards, synergies...
Problem: overfitting to MetaStone AI
Future: smart mutation (as humans do)
Human interaction