Explicabilidade em agentes artificiais deliberativos com o padrão publish/subscribe

Abstract

The agent paradigm is well suited to cyber-physical systems due to its distributed and autonomous characteristics. It is also suitable for information systems with distributed data, which require independent extraction, transformation and presentation. In scenarios that affect people’s lives or that are complex for users and developers, it is essential to provide explanations, whether for reasons of accountability, trust or debugging. This work proposes an explainability model for MASs (Multi-Agent Systems), enabling the generation of explanations in widely used paradigms, such as OOP (Object-Oriented Paradigm) and EDP (Event-Driven Paradigm). The model offers structures and mechanisms that decouple the explainability aspects from the agent’s behavior. The basis of the explanations is the execution history of each agent, which records changes in its internal state (such as changes in beliefs, desires and intentions) throughout the deliberation cycles. From this history, the model infers cause and effect relationships, forming the fundamental structure of the explanations. In OOP implementation, tracking the internal state can generate a very extensive history, and can also make the code complex. To overcome this problem, we opted for POE, which saves only changes in the internal state following the Publish/Subscribe pattern. This implementation was evaluated in two proof-of-concept scenarios: victim rescue and CRM (Customer Relationship Management). The evaluation compared the explanations produced with a final explanation, in natural language, already known, verifying whether the interpretation of the generated explanations leads to the same result. Although the model is exemplified in goal-oriented agents, it is generic enough to be applied in several systems that use POE. The approach, with its computational implementation, captures the essential elements to explain the agent’s decisions, fulfilling the main objective. However, the work does not cover the generation of explanations in natural language, since the final explanation may require additional interpretation by humans or other algorithms. The possible advantages of EDP over OOP are also not evaluated.