LLM-Based Policy Generation for Distributed Adaptive Systems

Abstract

In a multi-agent system, the actions of one or more agents may lead to undesirable results that may affect the entire system. As a result, it is important to have controls in place to govern the actions of these agents. Polices have commonly been used to establish constraints to regulate the actions that are permitted or prohibited within a system. However, it is challenging for policy authors to manually formulate and verify policies in complex multi-agent systems. In this work, we propose a framework that applies large language models (LLMs) to policy generation and management to simplify the work of humans in the loop. LLMs have permeated many aspects of human life in a short space of time. They provide powerful capabilities in natural language processing such as understanding input, generating output, and classifying data. Since the Ontology Web Language (OWL), which is the cornerstone of the semantic web, is widely used to describe declarative policies, we use a set of incremental experiments to demonstrate that pre-trained transformers can be used to generate ontology-driven polices, thus enabling practitioners to better manage adaptive systems.