DouRD: enhancing DouDizhu AI with role-differentiated modeling

Abstract

Imperfect information games impose greater demands on AI decision-making than perfect information settings, requiring models to infer hidden information, reason about opponent strategies, and dynamically optimize policies under uncertainty. In this study, we proposed a novel role-differentiated modeling approach within the deep Monte Carlo framework to enhance DouDizhu AI, the challenging three-player asymmetric imperfect information game. Our method incorporated attention mechanisms with role-specific adaptations to investigate their differential impacts on the landlord and peasant roles. Key findings demonstrate that: (1) the landlord and peasant roles require fundamentally distinct model architectures: experiments confirmed their functional independence; (2) attention mechanisms exhibit role-dependent effectiveness: CBAM significantly improved Peasant strategy execution, whereas SE and ECA offered moderate gains, while Self-Attention showed no enhancement; (3) surprisingly, applying attention mechanisms to the landlord role led to performance degradation, reinforcing the superiority of LSTM for
this role. These results highlight the importance of role-aware architecture design in the imperfect information game setting, and challenge the universal applicability of attention mechanisms.