Introduction
In the field of multi-agent systems (MAS), negotiation is traditionally viewed through two primary lenses: utilitarianism, which focuses on maximizing numeric payoffs, and game theory, which seeks stable equilibria. While effective in closed systems, these approaches often overlook the intrinsic "information value" of interactions. When an agent fails to fulfill a contract, the impact extends beyond the immediate loss; it provides critical data about the agent's reliability and the state of the environment.
This paper introduces the Information-Based Agency framework, a novel architecture that bridges the gap between game theory and information theory. By treating negotiation as a process of information exchange grounded in structured ontologies, the framework provides a principled method for quantifying trust, reputation, and reliability using the mathematics of entropy and probability distributions.
The Information-Based Agency Framework
The proposed framework addresses three critical dimensions of agency: the utilitarian (value of the deal), the informational (knowledge gained about the world and other agents), and the semantic (contextual relationships between concepts).
Dual-Language Architecture
To manage these dimensions, agents are structured with a dual-language system:
- Communication Language ($\mathcal{C}$): Used for illocutionary acts, such as making offers, requesting information, or issuing promises.
- Internal Probabilistic Logic ($\mathcal{L}$): Used to maintain and update a world model ($M^t$). This logic allows the agent to represent uncertainty rather than relying on binary "true/false" facts.
Learning via Minimum Relative Entropy
At the heart of the agent's learning mechanism is the principle of Minimum Relative Entropy, specifically Kullback-Leibler (KL) Divergence. When an agent receives new information—whether a message from a peer or an observation of the environment—it treats this information as a constraint on its world model. The agent then updates its internal probability distribution to a new state that satisfies these constraints while remaining as close as possible to its prior beliefs. This ensures that updates are conservative and mathematically grounded.
Semantic Grounding and Generalization
A unique feature of this framework is its use of ontological grounding. By utilizing a structured "is-a" hierarchy, agents can calculate semantic similarity between different tasks or objects. This allows for sophisticated generalization. For example, if an agent fails to deliver on a contract involving "apples," the framework allows the observer to slightly degrade trust in contracts involving "pears," as both are semantically related under the "fruit" category. This prevents the agent from having to relearn trust for every specific instance, facilitating "common sense" inferences even when data is sparse.
Quantifying Trust and Reliability
The framework formally distinguishes between Commitment (what an agent promises to do) and Enactment (what the agent actually does). The discrepancy between these two states is measured via entropy.
- Reliability: An empirical estimate of the difference between an agent's commitments and its actual enactments over time.
- Trust: The mathematical expectation of reliability regarding a specific class of future commitments.
- Reputation: The aggregation of these trust and reliability measures across a social context or multi-agent network.
By deriving these metrics from information theory rather than ad-hoc heuristic scores, the framework provides a more robust and predictable model for agent behavior in complex environments.
Conclusion and Future Work
The Information-Based Agency framework offers a rigorous path forward for automated negotiation. By integrating semantic ontologies with entropic learning, it allows agents to navigate the "gap" between rigid contracts and the messy reality of enactment.
However, challenges remain. The computational cost of calculating minimum relative entropy across complex distributions is significant. While the authors suggest approximations using the "Sim method," scalability to real-time, large-scale systems is an ongoing area of research. Additionally, the framework's reliance on subjective belief functions requires careful tuning of decay rates and initial weights. Future research will focus on applying this architecture to practical domains such as supply chain management and automated service-level agreement (SLA) negotiations.
Summary updated