Abstract
Motivated by Shannon's model and the recent rise of self-supervised "World Models" in AI, this paper proposes a Unified Intelligence-Communication (UIC) model. The framework describes both single agents and multi-agent systems (MAS) by modeling the environment as a generic communication channel. By unifying rule-based, layered, and world-model-based architectures, the UIC model provides a theoretical basis for analyzing multi-modal interactions—both explicit and implicit—within complex systems.
Introduction
Historically, information theory and artificial intelligence have developed along parallel but separate paths. Claude Shannon’s foundational work focused on the reliable transmission of data through noisy channels, while AI research focused on the internal processing required for intelligent action. This paper, the first in a three-part series, bridges this gap by reframing intelligent agents as communication transceivers and the environment as a complex, multi-modal channel.
The Problem: Bridging Bits and Behavior
While Shannon’s model excels at describing data movement, it does not inherently account for an agent that actively modifies its environment or absorbs energy. Conversely, many AI models abstract away the physical reality of information gathering. The UIC model addresses this by asking: Is the Shannon channel model sufficient to describe an agent that doesn't just receive waveforms, but also moves objects and interacts physically?
The Unified Intelligence-Communication (UIC) Model
The authors propose the UIC model by synthesizing classical communication theory with modern AI frameworks.
The Environment as a Channel
In this model, the environment is treated as a set of multi-modal channels. All physical interactions—whether electromagnetic, mechanical, or matter-based—are viewed as waves. This "Wave Hypothesis" allows researchers to apply Shannon’s Channel Capacity to quantify the maximum information any physical interaction can carry.
As shown in the model above, an agent's perception is represented by a set of receivers, while its actions are represented by a set of transmitters. This symmetry allows for a unified analysis of explicit communication (messages) and implicit communication (observing actions).
Learning Agent Architecture
The UIC model further integrates high-level agent architectures, such as Yann LeCun’s World Models and John Holland’s Complex Adaptive Systems, into a cohesive "Learning Agent" structure.
This architecture includes several key components:
- Critic: Evaluates performance against standards and provides feedback.
- Learning Element: Updates the agent's knowledge base and makes changes to the performance element.
- Performance Element: Executes actions based on current knowledge and experiments.
- Problem Generator: Drives exploration by identifying learning goals and suggesting experiments to find new experiences.
Key Insights
- Perception as Optimal Receiving: Sensors are modeled as receivers, meaning their efficiency can be bounded by Shannon Capacity. An optimal sensor extracts the maximum possible information from environmental waves.
- Action as Modulation: Every physical action is viewed as "modulating" the environment with information, effectively treating the world as a medium for communication.
- The Role of the Problem Generator: The authors highlight the Problem Generator as a critical component for "Strong AI," suggesting that limited internal representations provide the necessary motivation for environmental exploration.
Conclusion and Future Work
The UIC model provides a theoretical blueprint for designing intelligent systems by mapping agent components onto a communication-theoretic framework. The authors conclude that any interaction can be quantified by Shannon Channel Capacity and that intelligence can be described as the interaction between sets of waves.
While this first part focuses on conceptual modeling, subsequent parts of the treatise will provide quantitative case studies—including analyses of natural bee colonies and artificial multi-robot systems—to demonstrate the model's predictive power and its implications for the emergence of collective intelligence.
Summary updated