Design Patterns for Explainable Agents (XAg)

The ability to explain the behaviour of the AI systems is a key aspect of building trust, especially for autonomous agent systems - how does one trust an agent whose behaviour can not be explained? In this work, we advocate the use of design patterns for developing explainable-by-design agents (XAg), to ensure explainability is an integral feature of agent systems rather than an “add-on” feature. We present TriQPAN (Trigger, Query, Process, Action and Notify), a design pattern for XAg. TriQPAN can be used to explain behaviours of any agent architecture and we show how this can be done to explain decisions such as why the agent chose to pursue a particular goal, why or why didn’t the agent choose a particular plan to achieve a goal, and so on. We term these queries as direct queries. Our framework also supports temporal correlation queries such as asking a search and rescue drone, “which locations did you visit and why? ”. We implemented TriQPAN in the SARL agent language, built-in to the goal reasoning engine, affording developers XAg with minimal overhead. The implementation will be made available for public use. We describe that implementation and apply it to two case studies illustrating the explanations produced, in practice.

Authors

• Sebastian Rodriguez RMIT University
• John Thangarajah RMIT University
• Andrew Davey RMIT University

Keywords

• AOSE
• Design Patterns
• Explainable Agents
• EMAS