Only the Tens Count, but Trumps Decide: Formalising and Benchmarking Dahal Jeet, a Nepali Variant of Mendikot
This paper formally defines the previously undocumented Nepali card game Dahal Jeet and presents a comprehensive computational benchmark demonstrating that search-based agents outperform learning methods, while revealing that trump length is a stronger predictor of victory than scoring cards and that controlling for deal variance is essential to accurately measuring human-agent performance differences.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of artificial intelligence, researchers often test their creations by pitting them against each other in games. These games serve as laboratories where scientists can see how well a computer program learns, plans, and adapts to uncertainty. Some games, like chess, are played with perfect information, meaning every piece on the board is visible to both sides. Others, like poker or bridge, are games of imperfect information, where players must make decisions without knowing the full hand of their opponents. For decades, the standard approach to mastering these hidden-information games has been to teach computers through trial and error, a method known as reinforcement learning. The hope has been that by playing millions of hands, an artificial agent would eventually discover strategies that surpass human intuition. However, this assumption that learning is always the best path is being challenged by a new line of inquiry that asks whether a different kind of thinking—planning ahead based on what is known—might be more effective in certain specific situations.
This question takes center stage in a study of a card game called Dahal Jeet, played daily in the Terai-Madhesh region of Nepal. The game is a variant of a family known as Court Piece, played by four people in two fixed partnerships. The rules are simple but create a unique strategic landscape: players deal thirteen cards each, and a trump suit is determined by a random draw from a player's hand. The goal is to win tricks, but points are awarded in a very specific way. Only the four tens in the deck count toward the score. A team wins the hand if they capture three or more tens. If the tens are split evenly, with two for each side, the winner is decided by who captured more tricks. This creates a situation where the most important cards are few and far between, and the outcome often hinges on a single tie-breaking condition rather than a steady accumulation of points.
Before this research began, the game existed only in the oral tradition of the villages where it was played. There was no written rulebook, no digital version, and no scientific analysis of how it worked. The first step for the researchers was to document the game formally. They spent years observing players across seven districts, recording the rules as they were actually played, and creating a precise digital version of the game. This allowed them to build a computer engine that could play the game exactly as humans do, serving as a foundation for testing different artificial intelligence strategies.
The researchers then set out to see which type of artificial intelligence could play Dahal Jeet best. They built twenty-six different computer agents, ranging from simple programs that followed basic rules to complex systems that used machine learning to improve over time. They also included agents that used a technique called determinized search. This method works by imagining many possible versions of the hidden cards, planning the best move for each version, and then choosing the move that works best on average. The team pitted every agent against every other agent in thousands of matches, using a rigorous statistical method to ensure that the results were not just a fluke of luck.
The findings were surprising. The most successful agents were not the ones that learned through trial and error. Instead, the top performers were the ones that used search and planning. The strongest agent, which used a specific type of search algorithm, achieved a rating that placed it well above all the learning-based systems. In fact, none of the learning agents, no matter how they were trained or what rewards they were given, managed to beat a carefully tuned set of hand-written rules. The study showed that in this particular game, the ability to look ahead and calculate possibilities was far more powerful than the ability to learn from experience.
One of the most intriguing discoveries concerned how the computer "saw" the game. The researchers expected that giving the learning agents more information—such as knowing which cards had already been played or which suits were missing—would help them win. Instead, they found that the agents performed better when they were given less information. A simplified version of the game state, which stripped away many details, actually led to a higher win rate than a rich, detailed description. This suggests that for this specific game, too much data can confuse a learning system, while a simpler view allows it to focus on what truly matters.
The study also looked at how much of the game's outcome was due to skill versus luck. By analyzing thousands of hands, the researchers found that the cards a player was dealt explained nearly forty percent of the difference in who won. This high level of chance meant that simply comparing win rates between humans and computers could be misleading. To get a true measure of skill, the researchers developed a new method. They took the exact hands played by human participants and replayed them with a computer agent sitting in the human's seat. This allowed them to cancel out the luck of the deal and see how the human's decisions compared to the computer's decisions on the same cards.
When they used this paired method, the results changed. While the computer's overall win rate against the humans looked like a coin toss, the paired analysis revealed that the computer would have won significantly more hands if it had been playing the humans' cards. The humans lost roughly nine more hands out of every hundred than the computer would have on the same deals. This proved that the computer was indeed better, but that the advantage was small and easily hidden by the randomness of the card distribution.
The research concludes that for games where the scoring cards are rare and the objective is complex, planning ahead is a more reliable path to mastery than learning from experience. The study also highlights the importance of how we measure success in games of chance. Without controlling for the luck of the deal, we might miss the true difference in skill between a human and a machine. By documenting a traditional game and applying rigorous scientific testing, the researchers have not only created a benchmark for artificial intelligence but also preserved a piece of cultural heritage, showing that even in a game played for fun, there are deep lessons about how we think and how we learn.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.