🤖 AI

InfoChess: A Game of Adversarial Inference and a Laboratory for Quantifiable Information Control

This paper introduces InfoChess, a novel adversarial game designed as a testbed for studying multi-agent inference under partial observability by replacing piece capture with probabilistic scoring based on king location, and demonstrates the effectiveness of reinforcement learning agents alongside new information-theoretic metrics for analyzing strategic information control.

Original authors: Kieran A. Murphy

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kieran A. Murphy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a game of chess, but with a twist: nobody can capture any pieces. You can't eat the opponent's pawns, knights, or even their king. In fact, the only way to "win" isn't by checkmating anyone, but by guessing where the opponent's king is hiding.

This is InfoChess, a new game invented by researchers to study how intelligent agents (like AI) learn to spy on each other while hiding their own secrets.

Here is a breakdown of the paper using simple analogies:

1. The Game: A Game of "Hot or Cold"

In regular chess, the goal is to destroy the enemy. In InfoChess, the goal is information.

The Setup: It's played on a standard 8x8 board. The enemy King starts in one of four secret spots.
The Rules: You move your pieces. Some pieces act like flashlights (Rooks and Bishops), revealing a long line of squares. Others act like walls (Pawns), blocking the enemy's view.
The Catch: You never see the whole board. You only see what your pieces can "see."
The Score: At the end of every turn, you get points based on how confident you are about where the enemy King is. If you think there's a 90% chance the King is on square A, and the King is actually there, you get a high score. If you guess wrong, you get nothing.

The Analogy: Imagine playing "Battleship," but instead of sinking ships, you are trying to guess where the enemy Admiral is standing. Every time you move a piece, you either shine a light on a new area or put up a wall to hide your own Admiral.

2. The Players: From "Clueless" to "Master Spies"

The researchers didn't just play the game; they built a team of AI agents with different levels of "brainpower" to see how they would play.

The Random Player: Moves pieces like a toddler throwing darts. Guesses the King's location by rolling a dice.
The "Flashlight" Player (VisMax): Only cares about turning on as many lights as possible. It doesn't care where the lights point, just that they are on. It's like a person sweeping a room with a flashlight, hoping to accidentally trip over the King.
The "Detective" Player (BeliefMax): This agent builds a mental map. It remembers where the King was last turn and calculates where the King probably is now. It moves its pieces to check the most likely hiding spots.
The "Hiding" Player: This agent is smart about defense. It moves its own King to the darkest corner of the board to avoid being seen.
The "Super Spy" (Reinforcement Learning Agent): This is the AI that learned by playing thousands of games against itself and others. It didn't just follow rules; it discovered new strategies that the humans and simpler AIs didn't think of.

3. The Secret Weapon: "Entropy" (The Measure of Confusion)

The paper uses some fancy math words like "entropy" and "cross-entropy," but think of them as Confusion Scores.

Belief Entropy: How confused is the player? If they think the King could be anywhere, their confusion is high. If they are sure the King is in one spot, their confusion is low.
The Goal: A good player wants low confusion about the enemy (they know where the enemy is) but high confusion for the enemy (the enemy is totally lost about where the player is).

The researchers found that the "Super Spy" AI learned to manipulate these scores better than anyone else. It knew exactly when to shine a light and when to hide in the shadows to keep the opponent guessing.

4. What Did They Learn?

The study revealed three big things:

Modeling the Enemy Matters: The agents that tried to understand how the opponent thinks (the "Detective" and "Super Spy") won much more often than those that just moved randomly or blindly.
Offense and Defense are Linked: The best players didn't just look for the enemy; they actively hid their own King. The "Super Spy" learned to balance shining a light on the enemy while simultaneously building a wall around itself.
AI Discovers New Tactics: The AI found a weird pattern: it started moving its Rooks (usually long-range pieces) more often than its Bishops, a strategy the simpler "Detective" agents never used. This shows that when you give an AI a pure information goal, it finds creative ways to solve it that humans might miss.

Why Does This Matter?

You might ask, "Who cares about a weird chess game?"

The researchers say this is a laboratory. Real-world problems—like cybersecurity, military strategy, or even negotiating a business deal—are often about partial information. You don't know what the other person is thinking, and you don't want them to know what you are planning.

InfoChess strips away the messy parts of real life (like money or physical violence) and isolates the core problem: How do I learn the truth while keeping my own secrets safe? By studying how AI plays this game, we can build better systems for secure communication, autonomous drones, and strategic decision-making.

In short: InfoChess is a playground where AI learns the art of the "perfect bluff" and the "perfect guess," teaching us how to navigate a world where we never have the full picture.

1. Problem Statement

The paper addresses a fundamental challenge in multi-agent systems and game theory: isolating competitive information acquisition from other strategic objectives (such as material gain or territorial control).

The Limitation of Existing Games: In traditional partially observable games (e.g., Kriegspiel, Dark Chess, Stratego, Poker), information gathering is instrumental to a secondary goal (winning the game). This entanglement makes it difficult to study the specific mechanics of belief modeling, concealment, and information gain in isolation.
The Gap: There is a lack of controlled environments where the sole objective is the adversarial inference of an opponent's latent state, allowing for precise quantification of information dynamics without the confounding variable of "capturing pieces."

2. Methodology: The InfoChess Game

The authors introduce InfoChess, a symmetric adversarial game designed to elevate information acquisition to the primary objective.

Game Mechanics

Board & Pieces: Played on an $8 \times 8$ board with standard chess piece types (King, Rook, Bishop, Pawn) but no piece captures.
Movement: All pieces move one square in any direction.
Visibility Rules:
- Rooks & Bishops: Cast extended lines of sight (rays) similar to standard chess.
- Pawns: Act as obstacles, blocking opponent visibility rays.
- General: All pieces reveal their immediate surroundings.
Objective: Players do not checkmate the opponent. Instead, they score points based on the probability mass they assign to the opponent's King location at the end of every turn.
Turn Structure: Each turn consists of:
1. Moving a non-King piece.
2. Moving the King.
3. Making an inference (assigning probabilities to all squares) about the opponent's King location.
Game Duration: Fixed horizon of 25 turns per side to balance initial information acquisition and subsequent concealment phases.

Agent Hierarchy

To explore the strategy space, the authors define a hierarchy of agents with increasing levels of opponent modeling:

Random: Uniform random moves and uniform belief.
VisMax (V): Greedily maximizes newly visible squares (information gain) assuming a uniform belief; King moves randomly.
BeliefMax (B): Greedily maximizes expected information gain using a learned belief model of the opponent's King; King moves randomly.
HidingVisMax (HV): Uses VisMax for non-King moves but moves the King to minimize predicted opponent visibility (greedy concealment).
HidingBeliefMax (HB): Combines BeliefMax (non-King) and HidingVisMax (King) strategies.
RL Agent: A Reinforcement Learning agent trained via REINFORCE to maximize score differentials. It uses a frozen Transformer encoder (pre-trained for belief modeling) and an MLP head to score moves.

Information-Theoretic Characterizations

The paper introduces specific metrics to analyze gameplay dynamics:

Belief Entropy ( $H(q)$ ): Measures the agent's total epistemic uncertainty about the opponent's King location.
Oracle Cross Entropy ( $H(p, q)$ ): Measures calibration error by comparing the agent's belief ( $q$ ) against the true distribution ( $p$ ). It decomposes into the entropy of the true state and the KL divergence (mismatch).
Observer Cross Entropy: Measures calibration through the action-induced observation channel. Since the agent only observes whether the King is revealed or hidden (not the exact location if hidden), this metric evaluates how well the agent predicts the observation rather than the latent state, isolating miscalibration from irreducible uncertainty.

3. Key Contributions

Novel Game Design: InfoChess provides a symmetric, capture-free environment where information acquisition is the explicit, sole objective, removing material incentives that confound analysis in other games.
Agent Hierarchy & RL Benchmark: The authors establish a baseline of heuristic agents ranging from simple random policies to sophisticated agents with learned belief models and concealment strategies. They demonstrate that a Reinforcement Learning agent outperforms all heuristic baselines.
Quantifiable Metrics: The paper proposes a suite of information-theoretic measures (Belief Entropy, Oracle Cross Entropy, Observer Cross Entropy) to disentangle:
- Epistemic uncertainty.
- Calibration mismatch.
- Uncertainty induced by adversarial movement.
Strategic Insights: The study reveals that modest opponent modeling yields significant gains and that action-dependent observation channels fundamentally constrain what can be learned about latent states.

4. Results

Performance Hierarchy:
- Heuristics: Agents with learned belief models (BeliefMax) outperform those with uniform beliefs (VisMax). Agents that actively hide their King (HidingVisMax, HidingBeliefMax) significantly reduce the opponent's score.
- RL vs. Heuristics: The RL agent outperforms all heuristic agents. Interestingly, BeliefMax is the strongest heuristic competitor against the RL agent, suggesting the RL agent exploits the reduced stochasticity of hiding behaviors found in other agents.
Strategic Dynamics:
- Hiding Behavior: Agents that hide the King successfully increase the opponent's belief entropy and lower the opponent's score.
- Piece Usage: Heuristic agents (VisMax) move Pawns frequently to reveal local squares. In contrast, the RL agent and BeliefMax move Pawns less often and utilize Rooks more frequently than Bishops, indicating a shift in strategy toward long-range visibility control rather than immediate local fog clearing.
Metric Analysis:
- Calibration: Oracle cross entropy largely tracks belief entropy, indicating agents are well-calibrated.
- RL Adaptation: Observer cross entropy shows a mismatch in later game stages against the RL agent, suggesting the RL agent learns to move in ways that violate the predictions of static belief models.

5. Significance

Testbed for Multi-Agent Inference: InfoChess serves as a controlled laboratory for studying adversarial inference, concealment, and coordination under partial observability.
Decoupling Objectives: By removing piece capture, the game isolates the "information game," allowing researchers to study how agents balance exploration (gaining info) and exploitation (concealing info) without the noise of material trade-offs.
Theoretical Implications: The work highlights how action-dependent observation channels (where the agent's move determines what it can see) constrain the identifiability of latent states. This has broader implications for Partially Observable Markov Decision Processes (POMDPs) and multi-agent reinforcement learning (MARL).
Open Source: The authors release the environment, agents, and a public interface to encourage further research in this domain.

In summary, InfoChess transforms the abstract concept of "information gain" into a concrete, quantifiable game mechanic, providing a rigorous framework for analyzing how intelligent agents infer, conceal, and adapt to one another in uncertain environments.