Quantifying Multi-Agent Interaction via Mutual Information

24 concepts19 papers1 proposal

How do we know when two agents are truly interacting?

In multi-agent systems, traditional metrics often fall short. For instance, evaluating agents based solely on their final task reward can mask a lack of genuine cooperation; agents might simply exploit brittle, pre-coordinated conventions rather than actively responding to one another (Association for Artificial Intelligence 2026 et al., Tessera et al.). Similarly, spatial metrics like Euclidean distance fail to distinguish between two entities that are merely close to each other and those actively engaged in social negotiation (Tolstaya et al.).

To capture the true dynamics of coordination, a growing body of research turns to information theory. By treating agents' actions, states, and trajectories as random variables, we can use measures like Mutual Information (MI) to precisely quantify how much one agent's behavior influences or depends on another.

The Intuition: Interaction as Uncertainty Reduction

At its core, mutual information $I(X; Y)$ measures how much knowing one variable $X$ reduces our uncertainty about another variable $Y$. In a multi-agent setting, if two agents are operating completely independently, knowing Agent A's action gives us zero information about Agent B's action ($I(A_1; A_2) = 0$). If they are tightly coordinated, knowing what Agent A does drastically reduces the uncertainty of what Agent B will do.

Trendafilov et al. proposed using this task-independent metric to evaluate computer-mediated collaboration, showing that the statistical correlation between actions provides a deeper window into team dynamics than performance scores alone.

This intuition is particularly powerful in trajectory forecasting for autonomous driving. Tolstaya et al. and Ettinger et al. leverage MI to build conditional behavior prediction frameworks. Instead of assuming a vehicle will follow a fixed path, they measure how much the predicted trajectory of a neighbor changes when conditioned on the autonomous vehicle's planned actions. If the statistical divergence between the conditional and unconditional distributions is high, a meaningful interaction is occurring (Ettinger et al., Zhang et al.). This allows an autonomous system to prioritize a distant but highly interactive merging vehicle over a nearby, irrelevant parked car.

Isolating True Coordination via Conditioning

A common pitfall in measuring interaction is environmental confounding: two agents might look coordinated simply because they are reacting to the same external stimulus (e.g., two drivers hitting the brakes because a traffic light turns red).

To isolate genuine behavioral dependency, researchers condition the mutual information on the environmental state, calculating Conditional Mutual Information:

$$I(A_1; A_2 \mid S)$$

By conditioning on the state $S$, we subtract the information provided by the environment, leaving only the direct influence the agents have on each other. Bonjour et al. use this exact mechanism to detect covert collusion in multi-agent games. Even if colluding players obscure their payoffs, their conditional behavioral dependency betrays their hidden cooperation.

However, calculating pairwise mutual information between every single agent becomes computationally intractable as the collective grows. To solve this scalability bottleneck while respecting the Centralized Training with Decentralized Execution (CTDE) paradigm, Barez et al. proposed measuring the mutual information between the global state and the joint actions of all agents. This captures how the collective behavior as a whole depends on the environment state, providing a scalable proxy for multi-agent coordination without the combinatorial explosion of agent-to-agent tracking.

Operationalizing Influence for Learning and Robustness

Quantifying interaction does more than just diagnose behavior—it can actively guide how agents learn. In environments with sparse rewards, agents often struggle to discover that they need to work together (e.g., one agent standing on a switch to open a door for another). Wang et al. introduced intrinsic rewards based on mutual information to encourage exploration. By rewarding an agent when its actions significantly alter the state transitions of its teammates, the system naturally incentivizes agents to discover critical coordination bottlenecks.

Conversely, we can use these metrics to design systems that are intentionally less dependent on real-time coordination, making them robust to failures. For example:

  • Communication Loss: Karabag et al. utilize total correlation—an extension of mutual information to multiple variables—to quantify and minimize the intrinsic dependencies between agents' state-action processes. By minimizing this metric during planning, they synthesize policies that maintain high performance even if communication is completely severed.
  • Adversarial Teammates: Lee et al. developed an adversarial framework that uses mutual information to surgically disrupt coordination during training. By forcing agents to maintain performance even when their informational dependencies are broken, the agents learn highly robust interaction structures that survive sensor or teammate failures.

Diagnosing System Difficulty and Strategy

Finally, information-theoretic metrics reveal the underlying structural difficulty of multi-agent tasks. Allen et al. introduced the Influence Gap, which measures the asymmetry in how much different agents' actions inform environmental states and rewards. They discovered that highly asymmetrical systems (a large influence gap) are significantly easier for decentralized algorithms to solve, proving that problem difficulty is fundamentally tied to the distribution of control.

In a similar vein, Tessera et al. use mutual information between observations/hidden states and actions ($I(O; A)$ and $I(H; A)$) to audit whether reinforcement learning agents are truly grounding their decisions in environmental memory or merely coasting on brittle conventions. Meanwhile, Tian et al. use the Level of Influence (MI between an agent's reward and its partner's strategy selection) to identify exactly which training scenarios require diverse partner play, allowing for highly efficient allocation of computational resources.

Whether used to detect collusion (Bonjour et al.), incentivize truth-telling (Kong et al.), or map the emergence of collective behavior (Ferrarotti et al., Tuci et al.), information theory provides a unified, mathematical lens to understand, design, and optimize multi-agent interaction.

Key Foundational Work

Mutual Information as a Measure of Coordination in Collaborative Interaction (2015) — contributed 2 concept elements to this direction.
Agent Influence as a Predictor of Difficulty for Decentralized Problem-Solving (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_state_joint_action_mutual_information_ed81b578,
  title = {state-joint action mutual information},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/state-joint-action-mutual-information/rdr_ed81b578e14d41288a46faa9beed2816},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_state_joint_action_mutual_information_ed81b578_v1,
  title = {state-joint action mutual information},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/state-joint-action-mutual-information/rdr_ed81b578e14d41288a46faa9beed2816/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake