Research map / Multi-agent systems and games / Games and equilibria
DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning
Daochen Zha, Jingru Xie, Wenye Ma, Sheng Zhang, Xiangru Lian, Xia Hu, Ji Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DouZero's self-play reinforcement learning instead of supervised learning from expert matches | DouDizhu agents with the same neural architecture · DouDizhu vs SL | WP: 0.341 -> 0.659 | Improved |
| Use DouZero's Deep Monte-Carlo learning instead of DeltaDou's search-based approach | DouDizhu agents · 10,000-deck tournament vs DeltaDou | WP: 0.414 -> 0.586 | Improved |
| Train the card-playing policy without bidding in the learning loop, then evaluate it with a shared bidding network | DouZero · 100,000-deck tournament with bidding | WP: 0.461 -> 0.580 vs DeltaDou | Improved |
| Replace Deep Monte-Carlo returns with a Temporal-Difference objective (ablation) | DouZero · DouDizhu vs SL | WP | No clear effect |
| Replace Deep Monte-Carlo with an actor-critic using action features in the critic and masks in the actor (ablation) | DouZero · DouDizhu vs SL | WP | Hurt |
| Increase the number of parallel actors from 15 to 30 or 45 (ablation) | DouZero training · Training time against SL | WP | Improved |
| Use WP rather than ADP as the training reward (ablation) | DouZero · DouDizhu tournaments | WP | Improved |
| Use ADP rather than WP as the training reward (ablation) | DouZero · DouDizhu tournaments | ADP | Improved |
| Continue self-play training beyond five days | DouZero · Human expert data | accuracy | Hurt |
Later papers that compared with this method
1 later accepted paper reports DouZero as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- GAE Falls Short in Imperfect-Information Self-Play Reinforcement Learning (NeurIPS 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PerfectDou: Dominating DouDizhu with Perfect Information Distillation (NeurIPS 2022 · similarity 0.87)
- Simplified Action Decoder for Deep Multi-Agent Reinforcement Learning (ICLR 2020 · similarity 0.85)
- FightLadder: A Benchmark for Competitive Multi-Agent Reinforcement Learning (ICML 2024 · similarity 0.84)
- Jumanji: a Diverse Suite of Scalable Reinforcement Learning Environments in JAX (ICLR 2024 · similarity 0.83)
- BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games (ICLR 2025 · similarity 0.82)
- Can Large Language Models Master Complex Card Games? (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
