Implement smaller versions of games

[daochenzha]

Human-sized games could be too complex for the algorithms. We will implement smaller versions of the games like Dou Dizhu, Majong, UNO, to make it feasible for research. Thanks for the feedback from the anonymous reviewers.

[daochenzha]

We are brainstorming how we could develop smaller versions of the large games while capturing the key features of the game. For example, Majong is too complex since it is played by more than a hundred of cards and the winning conditions are too harsh, i.e., four melds and a pair. It would take a lot of time to train agents on this human-sized game.

A direction would be to first study small games, then proceed to large games. An example of smaller version of Majong could be using less cards and easier winning conditions, such as 1 meld and a pair. In this way, we keep the key features of Majong and make the game easier.

Any helpful comments about how we should implement smaller versions for games such as Dou Dizhu, Majong, and UNO, would be greatly appreciated :)

[DonaldTsang]

Here is a clue into what can be done for training common trick-taking games: Go for 32 card German or Piquet pack games first, since those have a smaller scale.
https://en.wikipedia.org/wiki/Template:Trick-taking_card_games

[DonaldTsang]

There are also many micro-state pokers out there https://wizardofodds.com/games/#poker-variants and of course blackjack https://wizardofodds.com/games/#blackjack-variants and Baccarat https://wizardofodds.com/games/#asian-games

[DonaldTsang]

If we are dealing with reduced state Mahjong then these historical variants might be useful

• https://en.wikipedia.org/wiki/Four_Color_Cards and https://en.wikipedia.org/wiki/Tam_c%C3%BAc which are Chinese chess based
• https://en.wikipedia.org/wiki/Chinese_dominoes which is very common in HK circles
• https://en.wikipedia.org/wiki/Khanhoo https://en.wikipedia.org/wiki/Madiao and https://en.wikipedia.org/wiki/Zi_pai which are all similar