Lu Yunlong from China wins the Sherlock Holmes Mahjong Mystery Solving Contest

The 5th season of “Sherlock Holmes and Mahjong” mystery solving contest is over. For two months 12 contestants from 6 countries have sent their solutions trying to solve nine mahjong mysteries.

The winner of the fifth Sherlock Mahjong Mysteries Contest is Lu Yunlong from China, AI researcher. Below one can find an interview with the winner.

An Interview

When and how did you learn to play mahjong?

Though mahjong is a very popular game played in China with many regional variants, I did not learn to play it until 2020 when I started my PHD in Peking University. Some seniors in my lab were doing research on AI application in various games. They created an online platform to host different games where users can submit AI agents to play these games and compete with each other. One of them implemented MCR on the platform and held a Mahjong AI competition in 2020. I found this game both intersting and challenging because of the variety of MCR fans. After I learned the basic rules of MCR, I played thousands of hands on the platform with the best AI agent in the competition and gradually improved my strategies by imitating its behavior.

What is your favourite style to play?

My favourite style to play MCR is a combination of much attack and a little defence. Attacking in MCR means to approach a winning hand as fast as possible, which involves choosing between different main fans with 6+ pts and (sometimes) a combinations of smaller fans. Sometimes I also give up attacking and make defence, when some other players are highly likely to make some fans (such as Little Three Dragons, All Types, Half Flush, etc.) given their melds and discarding zone, and I would destroy own hand to keep dangerous tiles in hand.

How did you discover “Sherlock Holmes and Mahjong” mysteries?

I discovered MahjongSoft platform in 2023 and started to play MCR on it. When exploring the website, I found and read the book “Sherlock Holmes and Mahjong” by Vitaly and was fascinated by the series of mysteries.

What are you feeling about participation in the current contest and winning it?

I feel quite happy to be able to participate in it and become the winner. I was already looking forward to participate in the next contest when I read the book in 2023, because I could not help but always read the answers before I solved them by myself. I was wondering how many questions I could solve on my own if the answers were not available. The contest this year gives me such a chance and it turns out that I can really solve mysteries in different levels of difficulty on my own. I think the contest mainly involves math and logical thinking combined with enough understanding of MCR rules. Solving the first few mysteries gave me confidence and I started to work on the next mysteries upon their release, which happened to give me advantages in the timestamp criterion.

Any ideas how to promote “Sherlock Holmes and Mahjong” mysteries in China?

I think the mysteries are intersting enough to attract many MCR players. However, lack of Chinese versions is the main obstacle from its promotion in China, because most of the MCR players in China cannot understand MCR terms like fan names even if they can read a little English. I may help to do the translation of these mysteries after I finish my Ph.D thesis and have more time in the next year, because they are really interesting and I want to share them to my friends who do not speak English.

Can you share more about your research topic of AI application in Mahjong?

Sure. Based on the online game AI platform developed by our lab, our team has hosted four Mahjong AI Competition in conjuction with IJCAI (a top-tier conference in AI) since 2020. We organized symposiums each year for the participants to share their methods and techniques in building their Mahjong AI agents. From their presentations we summarize three types of algorithms: heuristic methods which put human knowledge in the AI, supervised learning to train neural network models from datasets of gameplay, and reinforcement learning which can learn from scratch without human knowledge. I have published two papers in this topic, one [1] introducing MCR as a new testbed and benchmark of AI research, the other [2] summarizing the results of previous Mahjong AI competitions. We are also cooperating with another lab who have trained the strongest MCR AI agent ever using an innovative algorithm, which achieves superhuman performance in some initial tests. We are conducting large-scale experiments with some professional players to further test its capabilities and hopefully will publish another paper in a few months.

[1] Lu, Yunlong, Wenxin Li, and Wenlong Li. “Official International Mahjong: a New Playground for AI Research.” Algorithms 16.5 (2023): 235.

[2] Lu, Yunlong, Wenxin Li. Mahjong AI Competition: Exploring AI Application in Complex Real-World Games. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. 2024. Pages 8733-8736.