Abstract
The concept of coequal hand quality is proposed in this paper to judge the fairness while dealing cards in compound Doudizhu competitions, along with two coequal quality hand classifying solutions. The first one is to classify the hands according to mean values and standard deviation among scores of landlord-role players in gaming data, while the second one focuses on dimensionality reduced landlord-role player score probability distribution and a staged clustering algorithm. Experiment confirmed, the dimensionality reduced probability distribution based staged clustering classifying solution performs much better than the statistical solution.
Introduction
In competitive Doudizhu, a popular Chinese poker game, "compound competition" mechanisms are employed to minimize the element of luck. In these formats, different players are given identical hands to determine who performs better, thereby isolating pure skill. However, to maintain variety and engagement, organizers must generate different hands that are nonetheless of "coequal quality"—meaning they offer an equivalent level of advantage or disadvantage to the "landlord" player.
Defining and categorizing "hand quality" mathematically is a significant challenge. A hand's value is not a static number; its difficulty is represented by a complex distribution of potential outcomes. This paper addresses the limitations of simplistic statistical averages to ensure true fairness in high-stakes tournament settings.
Methodology
The research evaluates two distinct approaches to classifying hand quality:
1. Statistical Solution
This baseline method classifies hands using two primary metrics: the mean value (expected score) and standard deviation (volatility) of landlord-role player scores. Hands are divided into 10 categories based on these metrics, ranging from "landlord great inferiority" to "landlord great superiority," each further subdivided by volatility levels.
2. Staged Clustering Solution
This advanced approach treats hand quality as a probability distribution of scores.
- Feature Extraction: With 95 possible discrete scores, the authors reduced dimensionality into 7, 10, or 13 "sections" to facilitate machine learning.
- Clustering: The researchers utilized the Affinity Propagation (AP) algorithm followed by iterative hierarchical clustering to group similar hands.
- Evaluation: Cluster cohesion was measured using the inter-cluster average Kullback-Leibler (K-L) divergence (ckl). A lower ckl indicates that hands within a cluster are more truly "coequal."
Data Analysis and Insights
The study utilized a massive dataset from a major Chinese gaming platform, comprising approximately 1.8 million gaming records covering 320,000 unique hand patterns. After filtering for reliability (hands played at least 10 times), 12,000 patterns were selected for the final analysis.
A key observation from the data is that the frequency distribution of landlord scores approximates a Gaussian distribution. While mean and standard deviation provide a summary, they are "lossy" compressions. Probability distributions capture the "native and vivid" character of a hand, revealing that hands with identical averages can have vastly different win/loss probabilities.
Key Results
The staged clustering approach proved vastly superior to the traditional statistical method:
- Cohesion (ckl): The statistical solution resulted in a ckl of 1.871678, whereas the staged clustering solution achieved a significantly lower ckl of 0.5.
- Practical Application: The 10-dimension clustering model successfully grouped hands into 1,728 distinct clusters, providing a robust library of coequal hands for tournament organizers.
Conclusion and Future Work
While the clustering method offers a clear improvement in ensuring fairness, future research could address player skill bias. Currently, the data does not distinguish between expert and novice play, which may skew the perceived quality of certain hands. Additionally, exploring the computational scalability of Affinity Propagation for even larger datasets and adapting these methods for games with different hidden information structures, such as Texas Hold'em, remains a priority for future study.
Summary updated