Due to the recent success of machine (ML) and deep learning (DL) and the resurgence of other AI techniques, in the next decade, academia will focus on engineering intelligent systems, possibly involving data-driven, trainable components.
One of the critical challenges in intelligent system engineering is integrating diverse AI technologies while preserving conceptual integrity. Indeed, tomorrow's intelligent systems shall embody many innovative capacities – such as image, speech, and text recognition and generation – other than the criterion about when, how, and if to exploit those capabilities to support human users or cooperate with each other smoothly.
The most widely accepted solution to this challenge is using agents and Multi-Agent Systems (MAS), possibly embedding i) data-driven solutions to support their intelligent capabilities, as well as ii) complex automated reasoning/planning solutions to support their autonomous decision-making.
However, as multi-agent systems become more and more sophisticated, the capability of humans to understand their behavior – hence trusting them and accepting their support – becomes weaker. Therefore, the need for explainable intelligent systems – which are capable of motivating their behavior to the users with some degree of autonomy – is more compelling than ever.
There is a growing divide within the AI community between the success of sub-symbolic techniques (DL/ML) and the public's concern about the role of intelligent systems in society, which has led to the need for eXplainable Artificial Intelligence (XAI). XAI aims to make AI systems (more) transparent and accountable by providing algorithmic and software tools to ease users' understanding of AI systems.
MAS have certainly a role to play in pushing AI towards higher degrees of transparency, accountability, or explainability. Indeed, as the place where automated decision-making occurs, agents may require further smart capabilities aimed at explaining their behavior. In this sense, MAS themselves could be a subject for the field of XAI—and this is even more true if they wrap ML/DL-based solutions, which are inherently poorly understandable, or complex planning or deliberation strategies. However, MAS may also act as a tool for the field of XAI, as their social ability can be straightforwardly exploited to support any possible explanation process.
Finally, agents are where symbolic and sub-symbolic AI should meet. Arguably, explaining AI systems to humans should involve symbolic information—as symbols are what humans understand. Cognitive agents in intelligent MAS already use symbolic AI models and technologies for rational processes, knowledge representation, expressive communication, and effective coordination. The combination of MAS and symbolic AI has the potential to engineer explainable intelligent systems and has significant implications for fields such as robotics, computer science, and economics.
We welcome submissions from academia and industry that address the challenges and opportunities of XAI, from knowledge extraction (interpretability) to knowledge manipulation and sharing at an agent-agent and agent-human level. Finally, we encourage interdisciplinary submissions integrating perspectives from the computer science, software engineering, data science, legal AI, and ethics communities.
For the sake of reproducibility, submissions involving software contributions of any sort should include an URL to a publicly available repository (e.g., on GitHub) containing the source code and data/experiments if it is the case.
EXTRAAMAS 2023@AAMAS 2023
Autonomous Agents and Multi-Agent Systems (J.AAMAS)
From Large Language Models to Small Logic Programs: Building Global Explanations from Disagreeing Local Post-hoc Explainers (article in journal, 2024) — Andrea Agiollo, Luciano Cavalcante Siebert, Pradeep Kumar Murukannaiah, Andrea Omicini