Action Selection in Multiplayer Hidden-Information Games with Expected Target Entropy

Abstract

Hidden-information board games require agents to choose actions that reveal useful evidence under uncertainty. We study single-agent action selection under multiplayer observation mechanics in simplified environments through a reusable belief-state adapter and a one-step expected-target-entropy policy. The policy scores legal actions by how uncertain the agent expects to remain about a task target after acting, exposing information value across game-specific observation models. We evaluate simplified Clue, Sleuth, and Scotland Yard environments through systematic multi-seed experiments against random, heuristic, and observation-sampling baselines. Entropy selection reaches target identification in fewer turns in the two card games, while the hidden-movement pursuit task favors a distance heuristic. This contrast suggests that information objectives transfer most directly to target-identification games and should be paired with task-progress costs when control dominates. We extend expected-entropy action selection from single-agent deduction settings to multiplayer observation settings, motivating future agents that balance information gathering with task progress and control.

Citation

Xu, K., Meng, F., Verbrugge, C., & Lucas, S. (2026). Action Selection in Multiplayer Hidden-Information Games with Expected Target Entropy. In 2026 IEEE Conference on Games (CoG) (in press).

@inproceedings{xu2026expectedtargetentropy,
  author    = {Kaijie Xu and Fandi Meng and Clark Verbrugge and Simon Lucas},
  title     = {Action Selection in Multiplayer Hidden-Information Games with Expected Target Entropy},
  booktitle = {2026 IEEE Conference on Games (CoG)},
  year      = {2026},
  organization = {IEEE}
}