Human-Agent Cooperation in Games under Incomplete Information through Natural Language Communication

TL;DR

This work addresses cooperative planning between a human and an autonomous agent under incomplete information. It introduces a shared-control game built on Gnomes at Night and a language-driven pipeline that converts natural language into a compact flag representation to guide an asymmetric information-set Monte Carlo Tree Search (AISMCTS-F) planner. The approach combines a language module (LLM-based) with a planning module to produce flag-based policies and exchange, enabling more efficient human-agent cooperation, as demonstrated by human-subject experiments showing fewer turns and competitive completion times compared to a mute agent, though still short of human-human performance. Key contributions include the formulation of cooperative policy synthesis under private transitions, the AISMCTS-F algorithm with hidden information tracking, and empirical evidence that NL communication narrows information gaps in real-time collaboration. The results highlight the potential of natural language interfaces to enhance human-agent teamwork in settings with incomplete information and pave the way for richer intent representations and longer-horizon coordination.

Abstract

Developing autonomous agents that can strategize and cooperate with humans under information asymmetry is challenging without effective communication in natural language. We introduce a shared-control game, where two players collectively control a token in alternating turns to achieve a common objective under incomplete information. We formulate a policy synthesis problem for an autonomous agent in this game with a human as the other player. To solve this problem, we propose a communication-based approach comprising a language module and a planning module. The language module translates natural language messages into and from a finite set of flags, a compact representation defined to capture player intents. The planning module leverages these flags to compute a policy using an asymmetric information-set Monte Carlo tree search with flag exchange algorithm we present. We evaluate the effectiveness of this approach in a testbed based on Gnomes at Night, a search-and-find maze board game. Results of human subject experiments show that communication narrows the information gap between players and enhances human-agent cooperation efficiency with fewer turns.

Figures (10)

• Figure 1: An illustration of the shared-control game in \ref{['def:cooperative_control_game']}.
• Figure 2: Gnomes at Night testbed where the token is the magnetically connected gnome pieces and two private transition functions encode the wall layouts on each side of the board. Middle figure credit to gnomesAtNight2016.
• Figure 2: Summary of the number of gameplay instances collected for each type of ego player---human ($\textit{H}$), agent with communication ($\textit{A-Comm}$), and agent without communication ($\textit{A-Mute}$)---over five rounds.
• Figure 3: Communication-based approach that generates $\pi^\textbf{E}:\mathcal{S}\times M\to \mathcal{A}^\textbf{E}\times M$ via a language module (orange trapezoid) and a planning module (brown rectangle) interfaced through flags.
• Figure 4: AISMCTS-F: a tree from the perspective of an ego player $\textbf{E}$ in a minimal maze example shown at the top.

Theorems & Definitions (1)

• Definition 1