TLDR: A study on AI agents for the traditional Dhumbal card game found that a rule-based ‘Aggressive’ agent overwhelmingly outperformed sophisticated search-based (ISMCTS) and reinforcement learning (PPO) agents. The Aggressive agent achieved an 88.3% win rate by effectively exploiting the game’s ‘Jhyap’ declaration mechanic, highlighting the power of simple, game-specific heuristics in games with moderate information asymmetry.
A recent study has delved into the strategic depths of Dhumbal, a traditional card game popular in South Asia, by developing and comparing various Artificial Intelligence (AI) agents. Dhumbal, also known as Jhyap in Nepal and Yaniv in Israel, is a draw-and-discard game that involves strategic decision-making, managing imperfect information (players don’t know each other’s hands), and risk management. It’s a game that fosters social bonds and is culturally significant, making it an excellent candidate for AI research and digital preservation.
The objective in Dhumbal is to minimize the total point value of cards in hand. Players can discard single cards, sets of identical ranks, or sequences of three or more consecutive cards of the same suit. A key strategic element is the ‘Jhyap’ declaration: if a player’s hand value is 10 points or less at the start of their turn, they can declare ‘Jhyap,’ initiating a showdown where the lowest hand wins. This declaration carries risk, as a failed Jhyap can result in significant penalties.
Exploring Diverse AI Strategies
The research, conducted by Sahaj Raj Malla, aimed to identify the most effective AI strategy for Dhumbal and analyze how different AI paradigms handle the game’s unique challenges. The study implemented a diverse range of AI agents, categorized into four main types:
- Rule-Based Agents: These agents follow predefined heuristics or rules. Four variants were developed: Aggressive, Conservative, Balanced, and Opportunistic, each with a distinct risk profile.
- Search-Based Agents: These agents explore possible future game states to find optimal actions. Two methods were used: Monte Carlo Tree Search (MCTS) and Information Set Monte Carlo Tree Search (ISMCTS), which is designed to handle hidden information by sampling possible game states.
- Learning-Based Agents: These agents learn strategies through experience, typically via self-play. Deep Q-Network (DQN) and Proximal Policy Optimization (PPO), both reinforcement learning approaches, were implemented.
- Random Agent: A baseline agent that makes decisions purely by chance, providing a benchmark for evaluating the necessity of intelligent decision-making.
The agents were evaluated through a series of simulated tournaments, first within their categories and then in a cross-category championship. Performance was measured using metrics such as win rate, economic outcome (coin gains/losses), Jhyap success rate, and decision efficiency.
Unexpected Dominance: The Aggressive Agent
The results of the study presented a surprising outcome. While advanced search-based and learning-based methods have shown remarkable success in other complex games like Poker and Go, in Dhumbal, a simpler, rule-based approach proved overwhelmingly superior.
In the within-category tournaments:
- The Aggressive rule-based agent emerged as the strongest among its peers, demonstrating a superior win rate and positive economic performance.
- ISMCTS outperformed MCTS, highlighting the importance of handling imperfect information effectively in card games.
- PPO showed better performance than DQN in the learning-based category.
However, the true test came in the cross-category championship, where the Aggressive agent faced off against ISMCTS, PPO, and the Random agent. The Aggressive agent achieved an astounding 88.3% win rate, along with exceptional economic performance, consistently gaining coins. In stark contrast, ISMCTS managed a 9.0% win rate, while PPO and the Random agent performed poorly, with win rates of 1.5% and 1.3% respectively, and significant economic losses.
The Aggressive agent’s success is attributed to its effective exploitation of Dhumbal’s mechanics, particularly its strategy of frequent Jhyap declarations when its hand value is low. This allows it to end rounds quickly, often catching opponents with higher-value hands and incurring penalties. Its rapid decision-making (negligible decision time) also contributed to its efficiency, unlike the computationally intensive ISMCTS.
Also Read:
- When AI Judges Games: Fairness, Fun, and Human Alignment
- Unlocking Novice Minds: How We Reason About New Games Without Experience
Implications and Future Directions
This study suggests that for games like Dhumbal, with moderate information asymmetry and specific strategic thresholds (like the Jhyap declaration), carefully engineered heuristics can significantly outperform more complex, computationally demanding AI methods. This finding contrasts with the general trend in AI research for imperfect-information games, where advanced search and learning algorithms often dominate.
The research provides a reproducible AI framework and open-source code, contributing to both algorithmic research and the digital preservation of cultural games. Future work could explore the generalizability of the Aggressive agent’s strategy to other card games, evaluate its performance against human players, and investigate more sophisticated multi-agent reinforcement learning frameworks to improve the performance of learning-based agents in Dhumbal’s competitive environment.
For more details, you can read the full research paper here.
