Dou Dizhu, or Fighting the Landlord, is a shedding game involving 3 players and a deck of cards plus 2 jokers with suits being irrelevant. Heuristically, one player is designated the “Landlord” and the others become the “Peasants”. The objective of the game is to be the first one to have no cards left. If the first person to have no cards left is part of the “Peasant” team, then all members of the “Peasant” team receive a reward (+1). If the “Landlord” wins, then only the “Landlord” receives a reward (+1).

The “Landlord” plays first by putting down a combination of cards. The next player, may pass or put down a higher combination of cards that beat the previous play.

Our implementation wraps RLCard and you can refer to its documentation for additional details. Please cite their work if you use this game in research.

Arguments

dou_dizhu_v4.env(opponents_hand_visible=False)

opponents_hand_visible: Set to True to observe the entire observation space as described in Observation Space below. Setting it to False will remove any observation of the opponent’ hands.

Observation Space

The observation is a dictionary which contains an 'observation' element which is the usual RL observation described below, and an 'action_mask' which holds the legal moves, described in the Legal Actions Mask section.

The main Observation Space is encoded in a 1D vector with different concatenated features depending on the agent. To represent a combination of cards a 54-dimensional one-hot vector is encoded as follows. A 4x15 is constructed, where each column represents the rank of the cards (including the two jokers), and each row the number of matching card rank. The matrix is constructed using one-hot encoding. Since there are two jokers in the deck, the six entries that are always zero in the columns of the jokers are removed. Finally, to form the 54-dimensional vector, the one-hot matrix is flatten. The columns and rows of the 4x15 matrix are encoded as follows:

Encoding per Plane

The obsesrvation space of the landlord and the two peasants have different dimensions. The different agent observation spaces encode the following features:

Landlord Observation

Peasant Observation

Legal Actions Mask

The legal moves available to the current agent are found in the action_mask element of the dictionary observation. The action_mask is a binary vector where each index of the vector represents whether the action is legal or not. The action_mask will be all zeros for any agent except the one whose turn it is. Taking an illegal move ends the game with a reward of -1 for the illegally moving agent and a reward of 0 for all other agents.

Action Space

The raw size of the action space of Dou Dizhu is 27,472. As a reminder, suits are irrelevant in Dou Dizhu.

Rewards

We modified the reward structure compared to RLCard. Instead of rewarding 0 to the losing player, we assigned a -1 reward to the losing agent.

Version History

• v4: Upgrade to RLCard 1.0.3 (major changes to observation and action space) (1.11.0)
• v3: Fixed bug in arbitrary calls to observe() (1.8.0)
• v2: Bumped RLCard version, bug fixes, legal action mask in observation replaced illegal move list in infos (1.5.0)
• v1: Bumped RLCard version, fixed observation space, adopted new agent iteration scheme where all agents are iterated over after they are done (1.4.0)
• v0: Initial versions release (1.0.0)