Multiagent trust modeling for trusted AI and improved online social networks

20多年来，人工智能（AI）多智能体系统研究人员一直在探索如何最好地模拟智能体的可信度，这在电子商务等应用中选择最有价值的合作伙伴时很有价值。确定信誉最好的代理的方法考虑了根据过去的行为预测未来的可靠性；人们也在努力推断同行对其他代理人的建议是否值得信赖。最近，随着组织和个人对人工智能的关注越来越多，可信任的人工智能问题也成为人们关注的焦点：制定鼓励接受人工智能解决方案的战略，减轻可能出现的担忧和不信任。我们研究的中心前提是，多智能体信任建模方法可能是设计确保可信人工智能方法的关键。我们将说明缺乏信任建模如何阻碍一般的人工智能解决方案，以及校准基于信任的模型的方法如何有助于评估可信人工智能中各种努力的价值。将探讨三个关键子目标：映射执行特定于上下文的信任建模的解决方案；将信任建模应用于数字错误信息的关键社会问题展示信任如何能够为所有用户提供一个更易于访问和可接受的在线网络环境，以及理解意见动态在将这些社区聚集在一起方面所起的作用。在我们的工作中使用的一些主要方法将包括马尔可夫决策过程，用于从第一原则中逐步学习信任，从过去的行为中预测未来的可信度，整合以及用户建模；对仿真的信任建模方法进行了标定和比较，并对信任建模试验台的工作进行了研究；根据用户偏好调整解决方案，包括满足用户（老年人或有某种残疾的人）的辅助需求；基于网络动力学的社交网络影响者核心群体推理使用数据驱动的多方面方法，通过对作者和消息评分者进行建模来检测数字错误信息。这项工作将推进多智能体信任建模，整合贝叶斯和数据驱动方法，并强调信任的用途。这项工作的影响将是通过改进的信任建模方法，在人工智能解决方案中产生信任方面取得重大进展。在加拿大，投资于人工智能的好处的组织和个人将获得更好的信心，继续接受这些解决方案。受到可疑或压倒性内容挑战的社交媒体用户也将以更好的立足点出现。所有这些考虑都具有重要的社会价值；研究议程应该吸引目前代表性不足的群体的学生的重要参与，他们将因其在人工智能和社交网络方面的独特专业知识而受到追捧。

For over 20 years, artificial intelligence (AI) multiagent systems researchers have explored how best to model trustworthiness of agents, of value in selecting the most valuable partners for such applications as e-commerce. Methods for determining the most reputable agents have considered predicting future reliability based on past behaviour; there has also been effort in reasoning about the trustworthiness of peer advice about other agents. More recently, with increased attention on AI from organizations and individuals, the issue of trusted AI has come into focus as well: developing strategies for encouraging acceptance of AI solutions, allaying concerns and mistrust that may arise. The central premise of our research is that methods from multiagent trust modeling may hold the key in designing approaches for ensuring trusted AI. We will illustrate how lack of trust modeling handicaps general AI solutions and how the methods for calibrating trust-based models can be useful for assessing the value of varied efforts in trusted AI. Three key subgoals will be explored: mapping a solution for performing context-specific trust modeling; applying trust modeling to the critical social concern of digital misinformation; demonstrating how trust can serve to produce a more accessible and acceptable online networking environment for all users, understanding as well the role that opinion dynamics plays in bringing these communities together. Some primary methods used in our work will include Markov Decision Processes for progressively learning about trust from first principles, predicting future trustworthiness from past behaviour, integrating as well user modeling; simulations of trust modeling methods for calibrating and comparing, as is also done with the work on trust modeling testbeds; attuning solutions to user preferences including for assistive needs of users (older adults or those with some kind of impairment); reasoning about core groups of influencers in social networks through network dynamics; detecting digital misinformation through the modeling of authors and message raters, using a data-driven multi-faceted approach. This work will advance multiagent trust modeling, integrating Bayesian and data-driven methods and highlighting uses towards which trust is put. The impact of this work will be to enable significant steps forward with engendering trust in AI solutions through improved trust modeling methods. In Canada, organizations and individuals invested in seeing the benefits of AI will acquire better confidence in continuing to embrace these solutions. Users of social media challenged by questionable or overwhelming content will also emerge with a better footing. All of these considerations are ones with significant social value; the research agenda is one which should attract important participation from students of currently underrepresented groups, who will be sought for their unique expertise in AI and social networks.