# DouZero
DouZero is an open-source reinforcement learning agent for DouDizhu, the most popular card game in China[1](https://github.com/kwai/DouZero/), released in June 2021 by researchers at Kwai Inc. and [Texas A&M University](https://www.edgepedia.org/texas-a-and-m-university) and published at ICML 2021. According to its authors, it reached superhuman-level play using a deliberately simple method, Deep Monte-Carlo (DMC), without search, opponent modeling, or the heavy compute of the [AlphaZero](https://www.edgepedia.org/alphazero) lineage.
| Key fact | Detail |
|---|---|
| What it is | RL agent for DouDizhu, trained by self-play[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf) |
| Creators | Daochen Zha, Jingru Xie, Wenye Ma, Sheng Zhang, Xiangru Lian, Xia Hu, Ji Liu; Texas A&M University and Kwai Inc.[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf) |
| Released | Repository created 2 June 2021; ICML 2021 paper[1](https://github.com/kwai/DouZero/) |
| Method | Deep Monte-Carlo Q-learning, no search, and no opponent model[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf) |
| Training cost | Single server, four 1080 Ti GPUs, 48 processors; surpassed DeltaDou within 10 days (vendor-reported)[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf) |
| License | Apache 2.0[1](https://github.com/kwai/DouZero/) |
| Adoption | ~4,596 GitHub stars; integrated into RLCard; Botzone 2021 entry against 344 agents[1](https://github.com/kwai/DouZero/) • [3](https://www.douzero.org/) |
## What DouZero is
DouZero learns to play DouDizhu entirely from self-play. The ICML 2021 paper, "DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning," was authored by Daochen Zha, Jingru Xie, Wenye Ma, Sheng Zhang, Xiangru Lian, Xia Hu, and Ji Liu, affiliated with Texas A&M University and Kwai Inc. (Kuaishou), whose AI Platform team developed the framework[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf) • [1](https://github.com/kwai/DouZero/). The project's stated point is that classic Monte-Carlo methods can be made to deliver strong results in a hard domain with a complex action space[4](https://icml.cc/virtual/2021/poster/8445).
## Why DouDizhu is hard
DouDizhu is a three-player game in which one player, the Landlord, plays against two cooperating Peasants. The project documentation describes it as a challenging domain with competition, collaboration, imperfect information, a large state space, and particularly a massive set of possible actions where the legal actions vary significantly from turn to turn[1](https://github.com/kwai/DouZero/). The hidden hands mean the state cannot be fully observed; the two-against-one structure requires implicit teammate coordination; and the legal action set changes from turn to turn.
## The Deep Monte-Carlo method and architecture
**Deep Monte-Carlo.** The authors give three reasons DMC suits DouDizhu: it is not susceptible to overestimation bias when the action space is large; actions can be easily encoded as features rather than enumerated as outputs; and DMC parallelizes efficiently in wall-clock time[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf). Because actions are encoded as features by the network, the shifting legal-action set is handled without enumerating actions as outputs, and no search over opponent behavior is used.
**Networks and inputs.** The learner maintains three global Q-networks, one per seat (Landlord and the two Peasant positions), updated with MSE loss[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf). Each Q-network consists of an LSTM to encode historical moves and six MLP layers of hidden dimension 512[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf). Card combinations are encoded as one-hot 4×15 matrices, one row per count of a rank or joker, since suits are irrelevant in DouDizhu[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf); each matrix is flattened to a 54-dim one-hot vector[5](https://arxiv.org/abs/2106.06135). State features also include one-hot vectors of opponents' remaining cards (17 for a Peasant, 20 for the Landlord), a 15-dimension bomb count, and the most recent 15 moves encoded as a 5×162 matrix fed to the LSTM[5](https://arxiv.org/abs/2106.06135).
**Distributed setup.** Training used 45 parallel actor processes with one learner, communicating through three shared buffers[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf).
## By the numbers
All performance figures below are vendor-reported from the authors' own paper; no independent evaluation was found in the evidence base. Evaluation against existing DouDizhu programs used 10,000 randomly sampled decks, with algorithm A outperforming B if winning percentage (WP) exceeds 0.5 or average difference in points per game (ADP) exceeds 0. On this protocol DouZero dominated all rule-based and supervised-learning baselines and outperformed DeltaDou, then the strongest DouDizhu program[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf).
The compute claim is the paper's other headline: using a single server with four 1080 Ti GPUs and 48 processors, DouZero surpassed the supervised-learning baseline after 2 days of training and surpassed DeltaDou within 10 days[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf). The authors' poster summarizes the pipeline as requiring only days of training on 4 GPUs[6](https://dczha.com/files/poster-douzero-mastering-doudizhu.pdf).
On human alignment, the supervised-learning baseline, trained on 226,230 human expert matches filtered from the highest league of a popular DouDizhu mobile app, reached 84.2% move-prediction accuracy on held-out data. The authors report that DouZero aligned with human expertise in its first five days of training and then discovered novel strategies beyond human knowledge[2](https://proceedings.mlr.press/v139/zha21a/zha21a.pdf) • [5](https://arxiv.org/abs/2106.06135).
## Licensing, availability and reception
The code was released on GitHub on 2 June 2021 under the [Apache License](https://www.edgepedia.org/apache-license) 2.0[1](https://github.com/kwai/DouZero/). Pretrained checkpoints are hosted on [Google Drive](https://www.edgepedia.org/google-drive) and [Baidu Netdisk](https://www.edgepedia.org/baidu-netdisk) in two variants: DouZero-ADP, optimized for average difference in points, and DouZero-WP, optimized for winning percentage, alongside random, RLCard rule-based, and SL baselines for self-play evaluation[1](https://github.com/kwai/DouZero/). The toolkit was also integrated into RLCard[6](https://dczha.com/files/poster-douzero-mastering-doudizhu.pdf), making it a convenient baseline for card-game RL research. The repository showed about 4,596 stars and 649 forks at the retrieved snapshot[1](https://github.com/kwai/DouZero/), and the project site notes a Botzone 2021 competition entry against 344 AI agents[3](https://www.douzero.org/).
## References
1. kwai/DouZero GitHub repository. https://github.com/kwai/DouZero/
2. Zha et al., "DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning," ICML 2021 (PMLR v139). https://proceedings.mlr.press/v139/zha21a/zha21a.pdf
3. DouZero project website. https://www.douzero.org/
4. ICML 2021 virtual poster page for DouZero. https://icml.cc/virtual/2021/poster/8445
5. DouZero paper full text with appendices (arXiv 2106.06135), alphaXiv. https://www.alphaxiv.org/abs/2106.06135
6. ICML 2021 poster by the DouZero authors. https://dczha.com/files/poster-douzero-mastering-doudizhu.pdf
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Foundation-model methods and training › Reinforcement learning and world models*
*Initially written Sep 17, 2026 · Edited: Oct 11, 2026; Oct 15, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license