# AI solutions for drafting in Magic: the Gathering

**Authors:** Henry N. Ward, Daniel J. Brooks, Dan Troha, Bobby Mills, Arseny S. Khakhalin  
**arXiv:** [2009.00655](https://arxiv.org/abs/2009.00655)  
**Date:** 2020-09-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2009.00655v1/

## Key points

- Authors explore drafting in Magic: the Gathering, a complex sub-game
- Drafting involves building decks by selecting cards from a shared pool
- Challenges include complex mechanics, multiplayer dynamics, hidden information, and vast search space
- Lack of high-quality public datasets for drafting research
- Authors present dataset with over 100,000 simulated and anonymized human drafts from Draftsim.com
- Four distinct strategies for drafting agents: primitive heuristic agent, expert-tuned complex heuristic agent, Naive Bayes agent, deep neural network agent
- Deep neural network agent outperforms other agents in emulating human drafting behavior
- Naive Bayes and expert-tuned agents surpass simple heuristics in performance
- AI agents' accuracy analyzed across different stages of draft process and various cards to approximate subtle inconsistencies in human behavior
- Work contributes to development of more human-like drafting agents and provides benchmarks for future drafting bots through dataset provided

## Summary

In the paper "AI solutions for drafting in Magic: the Gathering," authors Henry N. Ward, Daniel J. Brooks, Dan Troha, Bobby Mills, and Arseny S. Khakhalin explore the sub-game of drafting in Magic: the Gathering, a popular trading card game. Drafting involves players progressively building decks by selecting cards from a shared pool. This process presents challenges for game-playing and AI research due to its complex mechanics, multiplayer dynamics, hidden information and vast search space. However, drafting has received limited attention in research due to the lack of high-quality public datasets. To address this gap, the authors present a dataset comprising over 100,000 simulated and anonymized human drafts collected from Draftsim.com. They also propose four distinct strategies for drafting agents: a primitive heuristic agent; an expert-tuned complex heuristic agent; a Naive Bayes agent; and a deep neural network agent. The authors benchmark these agents based on their ability to emulate human drafting behavior. The results demonstrate that the deep neural network agent outperforms all other agents in emulating human drafting patterns while both the Naive Bayes and expert-tuned agents surpass simple heuristics in performance. The accuracy of AI agents is analyzed across different stages of a draft process and for various cards to examine how well they approximate subtle inconsistencies in human behavior. By identifying unique strengths and weaknesses of each agent, this work contributes to advancing the development of more human-like drafting agents as well as providing valuable benchmarks for future generations of drafting bots through its dataset provided. Overall, this research sheds light on an understudied aspect of Magic: the Gathering gameplay and paves the way for further advancements in AI-driven drafting strategies.