Human-AI interactions in real-world complex uncertain environments using a comprehensive reinforcement learning framework

人工智能(AI)作为解决具体问题的核心技术方法的出现，是不可否认的，也是引人注目的。然而，人工智能仍然主要作为一个工具来执行狭隘的任务，而不是与人类用户关系中的支持合作伙伴。考虑到人类和人工智能各自的优势和劣势，这样的人机合作伙伴关系有可能超过其各部分的总和，利用互补的能力来实现原本仅凭其中之一是不可能或很难实现的结果。然而，为了使人工智能代理与人类用户/操作员尽可能协同工作，需要特定的方法、途径和技术。拟议研究的目的是推进这些技术和方法，并通过在复杂和敏感环境中的实际应用，在现实生活中展示它们的益处。该项目将把艾伯塔大学和雅各布中心的研究人员的努力与AIR和Thales提供的资源结合起来。在这次合作中，我们将研究如何利用人类知识在复杂和真实的环境中训练人工智能代理。它还将允许我们了解代理和人类的交互在何时何地允许我们实现比单独的代理或单独的人类更高的性能。为了实现这一点，将首先测试不同的框架、接口和协作类型。随后，反馈系统将被评估并相互比较，有时涉及人类和AI代理单独工作，有时涉及混合方法。预期的结果将使由AIR开发的COGMENT平台更广泛地被社区访问和使用。COGMENT平台是该项目使用的方法的核心。

The advent of Artificial Intelligence (AI) as a core technological approach to solve specific problems is both undeniable and remarkable. However, AI still operates primarily as a tool to execute narrow-focus tasks, rather than a supporting partner in a relationship with human users. Considering human and AI respective strengths and weaknesses, such a man-machine partnership has the potential to become more than the sum of its parts, leveraging complementary abilities to achieve results that would be otherwise impossible or very difficult to achieve with only one or the other. However, for AI agents to work as synergistically as possible with human users/operators, specific methods, approaches and technologies are warranted. The object of the proposed research is to advance those technologies and approaches, as well as to demonstrate their benefits in a real-life setting through practical application in a complex and sensitive environment. This project will combine the efforts of researchers from the University of Alberta and the JACOBB center with resources provided by AIR and Thales. Within this collaboration, we will investigate how human knowledge can be used to train AI agents in complex and real environments. It will also allow us to understand where and when the interaction of agents and humans allows us to achieve higher performance than agents alone or humans alone. To achieve this, different frameworks, interfaces and types of collaboration will first be tested. Subsequently, the feedback systems will be evaluated and compared with each other, sometimes involving humans and AI agents working alone, and sometimes a hybrid approach. The expected results will make the COGMENT platform, developed by AIR and at the heart of the methodology used for this project, more widely accessible and usable by the community.