[Paper Review] RedDwarfData: a simplified dataset of StarCraft matches

J. J. Merelo, Antonio Fernández-Ares+3 more|arXiv (Cornell University)|Dec 29, 2017

Artificial Intelligence in Games2 references3 citations

TL;DR

RedDwarfData presents a lightweight, pre-processed CSV dataset of 4,655 StarCraft matches derived from Robertson and Watson’s database, simplified by sampling resource changes rather than fixed frames. It enables efficient training and analysis of reinforcement learning and prediction models with balanced race distributions and ready-to-use temporal features, released under an open GitHub license for research and bot development.

ABSTRACT

The game Starcraft is one of the most interesting arenas to test new machine learning and computational intelligence techniques; however, StarCraft matches take a long time and creating a good dataset for training can be hard. Besides, analyzing match logs to extract the main characteristics can also be done in many different ways to the point that extracting and processing data itself can take an inordinate amount of time and of course, depending on what you choose, can bias learning algorithms. In this paper we present a simplified dataset extracted from the set of matches published by Robinson and Watson, which we have called RedDwarfData, containing several thousand matches processed to frames, so that temporal studies can also be undertaken. This dataset is available from GitHub under a free license. An initial analysis and appraisal of these matches is also made.

Motivation & Objective

• To address the challenge of high computational and data-processing overhead in using full StarCraft replay datasets for machine learning research.
• To create a simplified, accessible dataset that retains temporal dynamics while reducing resource demands for training and analysis.
• To provide a balanced, pre-processed dataset across all race combinations and match durations for use in supervised learning and bot evaluation.
• To enable faster experimentation in reinforcement learning and evolutionary algorithm-based bot development by offering a lightweight alternative to large raw replay datasets.
• To support early prediction of match outcomes and strategy analysis through structured, time-aligned feature extraction from game logs.

Proposed method

• Extracted data from Robertson and Watson’s original relational database of StarCraft matches, focusing on replay logs from TeamLiquid and BWReplays.
• Selected key events based on resource changes (minerals, vespene gas, supply) as time references, rather than fixed-frame sampling, to preserve temporal relevance.
• Aggregated unit counts and resource values across map regions by summing region-specific values and propagating the last known value forward to minimize missing data.
• Mapped enemy unit counts via estimated values based on visible units, acknowledging inherent uncertainty due to lack of visibility.
• Exported all processed data into standard CSV format for direct use in machine learning pipelines without additional parsing or execution.
• Ensured data balance by including matches across all race combinations (Zerg, Protoss, Terran) and durations, with no significant skew in distribution.

Experimental results

Research questions

• RQ1 Can a simplified, lightweight dataset derived from full StarCraft replay logs support effective machine learning training without sacrificing temporal and strategic fidelity?
• RQ2 How does the distribution of match duration and outcome vary across different race combinations in a large-scale, real-world replay dataset?
• RQ3 To what extent can resource change events serve as effective temporal anchors for modeling game progression in real-time strategy games?
• RQ4 Is the Zerg race inherently more dominant in terms of win rate across the dataset, and does this correlate with specific in-game dynamics?
• RQ5 Can a pre-processed, CSV-formatted dataset reduce the time and complexity of data preparation for reinforcement learning and bot evaluation?

Key findings

• The RedDwarfData dataset comprises 4,655 StarCraft matches with balanced representation across all race combinations, including 1,900+ matches in Terran vs. Protoss and significant Zerg participation.
• The average match duration is 25.26 minutes, with a mode between 20–25 minutes, indicating a relatively symmetric and non-skewed distribution.
• Zerg players won the highest proportion of matches they participated in, suggesting a potential strategic advantage in the dataset.
• Match duration is most affected by game dynamics rather than race identity, with Zerg vs. Zerg and Protoss vs. Protoss matches showing notable deviations in length.
• The dataset’s use of resource-change events as time references enables accurate temporal tracking while minimizing data bloat and missing values through last-known-value imputation.
• The dataset is available in open, standard CSV format on GitHub under an Apache license, enabling immediate use in machine learning and bot development workflows.