SCOUT: Capturing Human Strategic Intuition in "Race for the Galaxy" via Case-Based Reasoning

TL;DR

Building an AI that plays like a human—rather than just an optimized calculator—remains a major challenge in Game AI. SCOUT is an agent designed for the complex Eurogame Race for the Galaxy (RftG). By employing Case-Based Reasoning (CBR), SCOUT learns from a database of human wins to mimic strategic decisions. While it hasn't yet dethroned the top search-based engines, its 30% win rate and human-like play patterns mark a significant step for non-abstract game AI.

The Eurogame Challenge: Why RftG is Hard

Traditional board games like Chess are "solvable" through brute-force search and heuristics because they are deterministic and fully observable. Eurogames like Race for the Galaxy break this mold:

1. High Stochasticity: Cards are drawn randomly from a deck, requiring adaptivity.
2. Hidden Information: You don't know what's in your opponent's hand.
3. Complex Mechanics: Multiple paths to victory (Military, Trade, Production) create a massive decision space.

Previous agents, like the Keldon AI, use sophisticated search and domain-heavy evaluation functions. SCOUT takes a different path: What if we just look at what a human expert did in a similar situation?

Methodology: The CBR Loop

SCOUT's "brain" is divided into modular components (Phase, Placement, and Payment reasoning) coordinated by a Head module. The core logic relies on the 4R cycle of Case-Based Reasoning: Retrieve, Reuse, Revise, and Retain.

Architecture and Indexing

The agent uses a k-Nearest Neighbors (k-NN) algorithm to scan a "Case Base" of 1,000 games. To find the "most similar" situation, it checks 23 specific features, including:

• Player Tableau: The set of cards already played (Very High importance).
• Military Score: Crucial for specific expansion strategies.
• Hand Size: Indicates resources and flexibility.

Figure 1: The modular architecture of SCOUT, allowing for independent reasoning across game phases.

Experiments: Performance vs. Intuition

The researchers pitted SCOUT against the Keldon AI (an intermediate-to-advanced search-based agent).

The Quantitative Gap

While SCOUT doesn't beat Keldon AI consistently, the distance is smaller than it looks. The median score ratio was 0.85, meaning SCOUT usually stays within striking distance.

Mimicking Human Style

A key success for SCOUT was its Phase Selection. The frequency with which it chose actions like "Develop" or "Settle" mirrored the human expert much more closely than the Keldon AI did. It effectively learned the "Consume-Produce" strategy—a staple of high-level RftG play—simply by observing the expert's case-base.

Figure 2: Comparison of action selection frequencies between SCOUT, Keldon AI, and Humans.

Critical Insight: The "Why" behind the Results

SCOUT's biggest weakness is its lack of "Look-Ahead." Because it relies on historical cases (Retrieve/Reuse), it can sometimes fall for "plan-continuation" errors—making a decision because it felt right based on the past, even if the current draw makes it suboptimal.

However, its strength lies in Strategic Coherence. Unlike simple evaluation functions that might value cards in isolation, SCOUT recognizes synergies—it knows that certain cards belong together in a winning tableau because it saw a human do it.

Conclusion and Future Outlook

SCOUT demonstrates that CBR is a viable path for Eurogame AI, especially when domain knowledge is limited. The next frontier is clearly Hybridization. By combining SCOUT’s case-based strategic intuition with a Keldon-style search engine for tactical "nitty-gritty" (like exact resource math), we could create an agent that is not only stronger but also more human-like in its reasoning.

For developers, the takeaway is clear: If your game has a rich "meta" and complex card synergies, don't just build a calculator—build a librarian that remembers how the best players won.