# Evolutionary Approach to Collectible Card Game Arena Deckbuilding using Active Genes

arXiv: https://arxiv.org/abs/2001.01326 · HTML: https://mlsanity.com/paper/2001.01326/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 5 Jan 2020.

## Built on by (most cited first, 7 of 7)

- [Predicting Human Card Selection in Magic: The Gathering with Contextual Preference Ranking](https://mlsanity.com/paper/2105.11864/) (2021)
- [Summarizing Strategy Card Game AI Competition](https://mlsanity.com/paper/2305.11814/) (2023)
- [Evolving Evaluation Functions for Collectible Card Game AI](https://mlsanity.com/paper/2105.01115/) (2021)
- [UrzaGPT: LoRA-Tuned Large Language Models for Card Selection in Collectible Card Games](https://mlsanity.com/paper/2508.08382/) (2025)
- [Developing a Successful Bomberman Agent](https://mlsanity.com/paper/2203.09608/) (2022)
- [Two-Step Reinforcement Learning for Multistage Strategy Card Game](https://mlsanity.com/paper/2311.17305/) (2023)
- [A Taxonomy of Collectible Card Games from a Game-Playing AI Perspective](https://mlsanity.com/paper/2410.06299/) (2024)

## Abstract (arXiv)

In this paper, we evolve a card-choice strategy for the arena mode of Legends of Code and Magic, a programming game inspired by popular collectible card games like Hearthstone or TES: Legends. In the arena game mode, before each match, a player has to construct his deck choosing cards one by one from the previously unknown options. Such a scenario is difficult from the optimization point of view, as not only the fitness function is non-deterministic, but its value, even for a given problem instance, is impossible to be calculated directly and can only be estimated with simulation-based approaches. We propose a variant of the evolutionary algorithm that uses a concept of an active gene to reduce the range of the operators only to generation-specific subsequences of the genotype. Thus, we batched learning process and constrained evolutionary updates only to the cards relevant for the particular draft, without forgetting the knowledge from the previous tests. We developed and tested various implementations of this idea, investigating their performance by taking into account the computational cost of each variant. Performed experiments show that some of the introduced active-genes algorithms tend to learn faster and produce statistically better draft policies than the compared methods.