AI-Simulated Games of Machi Koro: 10,000 Game Turn Records

10,000 simulated games of the board game Machi Koro, generated by a neural network and reinforcement learning AI. Each row represents the game state at the beginning of a turn, with variables tracking player coins, constructed properties, and win conditions. The data was created to analyze game strategy and property usefulness.

Use Cases

• Train reinforcement learning agents for board games based on the simulated game history.
• Analyze the correlation between property acquisition and win rates based on the tracked property variables.
• Study player interaction dynamics, such as coin stealing and property switching, mentioned in the description.
• Benchmark game AI performance using the recorded win/loss outcomes.

Strengths

• Contains 10,000 complete simulated games, providing a substantial sample for analysis.
• Includes explicit win/loss labels and coin counts per player for outcome modeling.
• Tracks the construction of specific landmark properties (e.g., station, shopping mall) central to the game's objective.

Limitations

• Column-level documentation is absent; field semantics for most property variables must be inferred after download.
• Row count per game is unknown, which may limit suitability assessment for certain sequence models.
• Last update date is unknown; freshness unverified.

Provenance

Source

OpenML, sourced from a GitHub repository (mcandocia/machi_ai).

Collection Method

Generated via neural network and reinforcement learning simulations of the Machi Koro board game.