How A.I. Cracked an Ancient Roman Board Game

Researchers Used AI to Reconstruct How an Ancient Roman Game Was Played

More than a century after a scratched limestone slab was unearthed in the Netherlands, a new study in Antiquity reports a credible reconstruction of how the game on it may have been played. The team combined physical wear analysis with AI-driven simulations to test rule sets against the artifact's abrasion patterns.

The board is an oval block with a thin, dark rectangle incised on the surface-unusual enough that it puzzled historians for decades. Subtle grooves and polish visible from specific angles revealed repeat paths where hard pieces slid across stone. That was the clue.

The artifact, the place, the puzzle

The slab is held at the Het Romeins Museum in Heerlen, a city built over the Roman settlement Coriovallum. The museum focuses on local Roman life in Northern Europe during the first century.

Conservators have traced the incisions for documentation, showing a tight grid of lines on the surface and along the sides. Modern demonstration pieces sometimes shown with the stone were not found with the artifact.

Quick fact: Heerlen's Roman past

• Heerlen sits atop Coriovallum, a Roman village known for its baths-one of the best-preserved in the Netherlands.

The wear told the story

From a particular angle, the surface shows consistent polish along distinct lanes, with start-stop points at line intersections. That pattern indicates repeated sliding moves rather than random scratches or tool marks.

These constraints-where moves likely started, where they tended to end, and how often paths overlapped-became the ground truth the team tried to match with plausible rules.

How AI likely helped infer the rules

• Digitize the board's geometry and wear density: capture 2D/3D surface data under raking light to map abrasion intensity and directions.
• Generate candidate rule sets: define move types (slides, jumps, captures), legal zones, and goals consistent with Roman-era game traditions.
• Run large batches of simulated games with AI agents following each rule set to produce synthetic "wear" distributions.
• Score fit: compare simulated path frequencies and endpoints to the real wear map; iterate to improve alignment.
• Select a small set of rule configurations that best explain the observed wear without overfitting.

The outcome is a constrained family of ways to play that reproduces the board's observed traffic patterns. While not a definitive rulebook, it narrows the space from "mystery" to a testable, repeatable model.

Why this matters for science and research teams

• Method transfer: The same workflow-surface wear mapping + simulation-based inference-applies to tools, pathways, and instrument panels where user behavior leaves trace patterns.
• Evidence weighting: Physical constraints (abrasion) anchor models, while AI explores hypothesis space at scale. This balances empirical data and computational search.
• Reproducibility: Simulated traces and scoring functions make arguments auditable instead of purely interpretive.

Practical workflow you can reuse

• Capture: Raking-light photography and, where possible, structured-light or photogrammetric scans to get a detailed surface model.
• Preprocess: Normalize images, enhance micro-polish, and segment incisions without altering primary signal.
• Constrain: Convert visible wear to probabilistic paths (entry/exit points, intensity along lanes).
• Simulate: Enumerate rule sets; run many simulated sessions with stochastic players; log path densities.
• Compare: Use straightforward metrics (e.g., KL divergence or correlation of heatmaps) to score fit.
• Report: Publish top rule sets, parameters, and full comparison metrics so others can re-run or challenge the findings.

Limits and next steps

• Ambiguity remains: Multiple rule sets can produce similar wear. Report uncertainty, not a single "correct" answer.
• Confounds: Post-depositional scratches or modern handling can bias patterns. Document conservation history and exclude damaged zones.
• Cross-validation: Test the inferred rules on other boards of similar design, if available, to see if they produce consistent wear predictions.
• Experimental archaeology: Recreate the board and play under controlled conditions to see if fresh wear converges with the historical pattern.

Interested in applying these methods?

• Explore training and tools on applying AI to empirical research: AI for Science & Research
• Background on Roman games and material culture: Antiquity (journal)
• Institutional context for the research team: Leiden University - Archaeology

Bottom line

By treating surface wear as data and testing thousands of possibilities with AI, researchers transformed a puzzling stone into a playable model of Roman pastime. The approach is simple in spirit: let the artifact's scars set the rules, then let computation do the grinding work of elimination.