Multi-Personalities Guided Deep Monte Carlo Search for Complex Card Games: A Guandan Case Study

ICLR 2026 Conference Submission12785 Authors

Keywords: Monte Carlo Tree Search, Multi-agent Game, Human-like AI

Abstract: Human personality-enabled AI has become more and more important in many areas such as complex card games. In these complex multi-agent interaction scenarios, players' decisions are often significantly affected by personality and strategy style, which makes it difficult for traditional deep Monte Carlo methods to meet practical needs. In this paper, we propose a multi-personality guided deep Monte Carlo search framework, in which three rule-based personalities are incorporated as priors to bias the policy search. Experimental results show that the framework performs well in the Guandan game, which can quickly adapt to the rule prior and gradually discover better strategies through training. This study provides an effective solution for personalized decision-making in complex card games and multi-agent systems, and opens up a new direction for incorporating human style into deep reinforcement learning.

Primary Area: reinforcement learning

Submission Number: 12785