🔢 mathematics

Goal-Oriented Semantic Communication for Logical Decision Making

This paper proposes an explainable, goal-oriented semantic communication framework grounded in First-Order Logic that selects and transmits only the most critical logical clauses for collaborative decision-making in safety-critical environments, thereby optimizing bandwidth usage while ensuring transparency and verifiability.

Original authors: Ahmet Faruk Saz, Faramarz Fekri

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmet Faruk Saz, Faramarz Fekri

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 you are driving a self-driving car in a busy city. You have sensors that see everything: every pedestrian, every other car, every traffic light, and every pothole. Now, imagine you need to tell a central computer (or another car) what you see so it can help you make a safe decision, like "Stop!" or "Turn Left!"

The Problem: The Data Tsunami
If you tried to send everything your sensors see, you would be flooding the network with millions of tiny details. It's like trying to describe a whole movie by sending every single pixel of every frame. The network would get clogged, the message would arrive too late, and the decision would be made too slowly. In a safety-critical situation, that delay could be fatal.

The Old Way: The "Black Box" Guess
Current AI systems often try to solve this by compressing data using complex "black box" neural networks. They send a compressed version of the video, hoping the receiver can guess what's important. The problem? We don't really know why they chose what they sent. If the system makes a mistake, we can't easily check the math to see if it was logical. It's like a magician pulling a rabbit out of a hat; it works, but you can't explain the trick.

The New Way: The "Smart Detective"
This paper proposes a new way to communicate, called Goal-Oriented Semantic Communication. Instead of sending raw data, the car acts like a smart detective who only sends the clues that matter for the specific job at hand.

Here is how it works, broken down with simple analogies:

1. The Language of Logic (The Rulebook)

The system doesn't speak in pixels or sound waves; it speaks in First-Order Logic (FOL). Think of this as a universal rulebook for the city.

Instead of sending a photo of a red ball, it sends a sentence: "There is a pedestrian at the intersection."
It has a set of rules (like traffic laws): "IF a pedestrian is at the intersection, THEN stop."

2. The Goal: What Matters?

The car doesn't care about the color of the sky or the brand of a parked car. It only cares about Goal-Oriented States.

Analogy: Imagine you are playing a game of "20 Questions" to guess a secret word. You don't ask about the color of the word; you ask questions that eliminate half the possibilities at once.
In this system, the "Goal" is the decision (e.g., "Stop"). The system asks: "Which single piece of information, if I tell you, will make you 100% sure you need to stop?"

3. The Selection Process: The "Information Bottleneck"

The system uses a mathematical principle called the Semantic Information Bottleneck.

The Metaphor: Imagine you have a bucket full of water (all the data you see) and a tiny straw (the limited bandwidth). You need to suck up the water to put out a fire (make a decision).
Most people would just suck up random water.
This system is smart: it sucks up only the water that is actually on fire. It filters out the "noise" and only transmits the "signal" that changes the decision.

4. Why It's "Explainable" (The Verifiable Truth)

This is the paper's biggest breakthrough. Because the system uses logic rules (like math) instead of a black-box AI guess, it can prove why it made a choice.

Analogy: If a black-box AI says "Stop," you have to trust it. If this system says "Stop," it hands you a receipt: "I am stopping because I sent you the fact 'Pedestrian at Intersection,' and the rulebook says 'Pedestrian = Stop'."
This makes it perfect for safety-critical jobs like autonomous driving, where you need to know why a decision was made.

5. The Experiment: The "LogiCity"

The authors tested this in a simulated city called LogiCity.

They had cars driving around with limited communication.
They compared their "Smart Detective" method against a method that just picked random facts to send.
The Result: The "Smart Detective" was vastly superior. Even when allowed to send only one tiny fact, it made the right decision almost 100% of the time. The random method needed to send five times as much data to get the same result.

Summary

Think of this paper as a new way for robots to talk to each other. Instead of shouting a million words to describe a scene, they whisper the one specific sentence that changes the outcome.

Old Way: "Here is a 4K video of the street, good luck figuring out what to do."
New Way: "Pedestrian crossing. Stop."

It's faster, uses less data, and because it's based on logic, we can trust it to be safe and explainable. It turns communication from a data dump into a strategic conversation.

1. Problem Statement

In safety-critical autonomous systems (e.g., autonomous driving), agents often face a "bandwidth bottleneck" where transmitting raw sensory data is infeasible due to limited communication capacity. While existing semantic communication approaches exist, they generally fall into two categories with significant limitations:

Black-box Deep Learning: Methods using transformers or LLMs achieve compression but lack transparency, making it impossible to verify why specific data was selected or to guarantee safety-critical decisions.
Mathematical Information Theory: Existing theoretical frameworks offer guarantees but often lack a unified, explainable foundation tailored specifically for logical decision-making (deduction).

The core challenge is to develop a framework that:

Selects only the most task-relevant information (semantic content) to transmit.
Ensures the selection process is explainable and verifiable via formal logic.
Optimizes transmission to maximize the receiver's ability to make correct logical decisions (e.g., "Stop," "Go") rather than just reconstructing raw data.

2. Methodology

The authors propose a framework grounded in First-Order Logic (FOL) and Inductive Logical Probability.

A. Logical Representation of the World

The environment is modeled using a First-Order Language $\mathcal{L}$ with predicates (monadic and dyadic).

Constituents: The fundamental semantic units are "constituents" ( $C_{w,k}$ ), which are maximally specific descriptions of the world state (analogous to truth tables in propositional logic but for relational structures).
Inductive Probability: Instead of statistical frequency, the paper uses inductive logical probability (based on Carnap and Hintikka). The probability of a hypothesis is defined by the ratio of constituents compatible with the evidence to the total number of constituents.
- $c(\phi | e) = \frac{|C(e \land \phi)|}{|C(e)|}$
- This allows for a rigorous definition of Content Information: $cont(\phi) = 1 - c(\phi)$ , quantifying how much a statement eliminates possible world states.

B. Semantic Information Metrics

The authors define semantic analogs to Shannon information theory:

Semantic Entropy ( $H_s$ ): The average uncertainty of a set of hypotheses.
Semantic Mutual Information ( $I_s$ ): The reduction in uncertainty about a set of hypotheses given evidence.
Goal-Oriented States: Task rules (e.g., traffic laws) induce logical equivalence classes over world states. These are defined as Goal-Oriented States ( $\Phi$ ). The objective shifts from recovering the exact world state to resolving the truth value of these specific goal states.

C. The Optimization Objective

The problem is formulated as a Semantic Information Bottleneck:
$\max_{\hat{e}} I_s(\Phi; \hat{e}) - \beta I(e; \hat{e})$
Where $\hat{e}$ is the subset of evidence transmitted, $\Phi$ is the goal-state space, and $I(e; \hat{e})$ represents the transmission rate (cost).

Goal: Maximize the semantic mutual information between the transmitted evidence and the goal states while minimizing the amount of data sent.
Algorithmic Solution: The authors derive a polynomial-time lexicographical sorting algorithm.
- They show that under their inductive system, the objective function simplifies to minimizing a weighted sum of Bernoulli variances of the goal states.
- Due to the astronomical size of the constituent space, direct computation is impossible. The authors derive an asymptotic reduction where the objective is dominated by the smallest exponent in a specific formula.
- This leads to a symbolic comparison key $\kappa(\hat{e}) = (|I_{\sim}|, K, -H_{(1)}, -H_{(2)}, \dots)$ $κ (\overset{e}{^}) = (∣ I_{\sim} ∣, K, - H_{(1)}, - H_{(2)}, \dots)$ , where:
  - $K$ : Number of distinct Q-sentences (Evidence Diversity).
  - $H_i$ : Specificity of non-overlapping hypotheses.
- The algorithm selects the subset of $k$ FOL clauses that lexicographically minimizes this key, effectively prioritizing evidence that resolves the most uncertain goal states first.

3. Key Contributions

Principled Semantic Framework: A unified theory defining semantic information, entropy, and mutual information strictly within the context of FOL and inductive probability, bridging logic and information theory.
Goal-Oriented Formulation: Extending semantic communication from "state recovery" to "decision support" by defining goal-oriented states as logical equivalence classes induced by task rules.
Explainable Selection: Unlike deep learning "black boxes," the proposed selection mechanism is based on explicit logical clauses and mathematical proofs, ensuring verifiability.
Efficient Optimization Algorithm: The derivation of a lexicographic sorting algorithm that bypasses the need to compute astronomically large probability values, making the optimization tractable in polynomial time.
Empirical Validation: Demonstration in a neuro-symbolic urban simulator (LogiCity) showing significant performance gains over random selection.

4. Experimental Results

The framework was evaluated in the LogiCity simulator (a 241x241 grid with cars and pedestrians) under various rule sets and communication budgets.

Setup: Agents communicate via an edge server (GNA/LNA) to select $k$ FOL expressions to transmit.
Metrics:
- H-DSR: Hypothesis Decision Success Rate (agreement with full-information baseline).
- A-DSR: Action Decision Success Rate (correct action taken: Stop, Slow, Fast, Normal).
Key Findings:
- Low-Rate Dominance: Semantic selection vastly outperforms random selection, especially under tight bandwidth constraints.
  - Example: With $k=1$ (transmitting only 1 entity), semantic selection achieved an A-DSR of 0.984 (on spatial rules), whereas random selection was only 0.869.
  - Semantic selection reached near-perfect performance ( $A-DSR \approx 1.0$ ) by $k=3$ , while random selection required $k=5$ to reach only 0.952.
- Bandwidth Efficiency: Semantic selection at $k=1$ matched or exceeded random selection at $k=5$ , representing a 5x bandwidth saving.
- Robustness: The advantage was most pronounced in difficult scenarios (small Field of View, high pedestrian density), where the decision problem is hardest.
- Architecture: The method worked effectively across Sensor Global, Single-Zone, and Multi-Zone architectures, though gains were slightly lower in distributed settings due to the limited pool of available observations.

5. Significance

This paper represents a significant step toward AI-native, safety-critical communication systems.

Safety & Verifiability: By grounding communication in formal logic, the system provides a mathematical guarantee that the transmitted information is sufficient for specific logical deductions. This is crucial for regulatory approval in autonomous driving.
Efficiency: It proves that "meaning" can be compressed far more efficiently than raw data or random samples when the receiver's task is known.
Theoretical Bridge: It successfully unifies the fields of Inductive Logic, Semantic Information Theory, and Communication Engineering, offering a new paradigm where communication is optimized for decision quality rather than data fidelity.

In summary, the authors demonstrate that by using logical structures to define "what matters," autonomous agents can communicate with extreme efficiency while maintaining the transparency and safety required for real-world deployment.