Using Abstraction in Reasoning about Autonomous Agents and Multiagent Systems

当开发在复杂动态环境中执行任务的自主代理时，在规划代理行为、解释代理行为和进行强化学习时，抽象的使用是至关重要的。我们可以使用简化的抽象模型有效地生成高级解决方案，然后使用详细的具体模型对这些解决方案进行细化。抽象模型可以用人类理解的术语来表达，而具体模型可以被机器使用。更一般地说，我们可能有一个多层表示，其中使用各种模型在不同的细节级别上执行推理并处理不同类型的偶然事件。在最近与Banihashemi和De Giacomo的合作中，我开发了情境演算中主体抽象的正式描述，这是一个众所周知的用于推理行为的谓词逻辑框架。我们假设我们有一个高层次规范和一个低层次规范的代理，都表示为行动理论在情景演算。细化映射指定了每个高级操作如何由低级ConGolog程序实现，以及如何将每个高级谓词转换为低级公式。我们在这些行动理论之间定义了健全/完全抽象的概念。我们证明了健全/完整的抽象具有许多有用的属性，这些属性确保我们可以在抽象级别上推断代理的行为（例如，综合计划），然后在较低级别上改进得到的解决方案。该框架还可以用于生成低级行为的高级解释。在这个项目中，我和我的学生将扩展这项工作并将其应用于新的问题。首先，我们将研究如何将我们的描述推广到不确定性领域，在这些领域中，行为有许多不受代理控制的可能结果。其次，我们将研究如何利用抽象来执行可解释的计划，在这里我们弥合了人类用户模型和系统模型之间的差距。第三，我们将研究如何在“实际推理”中适应暂时扩展的目标和抽象计划，在“实际推理”中，智能体随着时间的推移逐步完善/修改她的意图，同时保持它们的一致性；代理不需要考虑完全详细的计划，应该能够利用关于目标/计划如何相互作用的知识。第四，我们将研究如何将抽象用于对其他主体的推理，以及它们如何帮助/干扰主体目标的实现；这将产生一种分散的多智能体认知规划形式，其中每个智能体生成自己的计划，将子目标委托给其他智能体，并充分了解其他智能体的能力、意图和合作意愿，以确信自己的目标将被实现。最后，我们将研究如何合成对给定应用程序/目的有用的抽象。

When developing autonomous agents that perform tasks in complex dynamic environments, the use of abstraction is crucial in planning agent action, in explaining the agent's behaviour, and in doing reinforcement learning. We can use a simplified abstract model to generate high-level solutions efficiently, and later refine these using the detailed concrete model. The abstract model may be expressed in terms that humans understand, while the concrete model can be used by the machine. More generally, we may have a multi-tier representation where various models are used to perform reasoning at different levels of detail and address different kinds of contingencies. In recent work with Banihashemi and De Giacomo, I have developed a formal account of agent abstraction in the situation calculus, a well known predicate logic framework for reasoning about action. We assume that we have a high--level specification and a low--level specification of the agent, both represented as action theories in the situation calculus. A refinement mapping specifies how each high--level action is implemented by a low--level ConGolog program and how each high--level predicate can be translated into a low--level formula. We define notions of sound/complete abstractions between such action theories. We showed that sound/complete abstractions have many useful properties that ensure that we can reason about the agent's actions (e.g., synthesize plans) at the abstract level, and then refine the solutions obtained at the low level. The framework can also be used to generate high--level explanations of low--level behavior. In this project, my students and I will extend this work and apply it to new problems. First, we will examine how to generalize our account to apply to nondeterministic domains, where actions have many possible outcomes that are not under the agent's control. Second, we will study how abstraction can be exploited to perform explainable planning, where we bridge the gap between the human user's model and the system's model. Third, we will examine how to accommodate temporally extended goals and abstract plans in "practical reasoning" where an agent progressively refines/revises her intentions over time, while keeping them consistent; the agent should not have to consider fully detailed plans and should be able to exploit knowledge about how goals/plans interact. Fourth, we will study how abstraction can be used in reasoning about other agents and how they can help/interfere with the accomplishment of the one's goals; this should yield a form of decentralized multi-agent epistemic planning, where each agent generates her own plan, delegating subgoals to other agents, and knows enough about the other agents' abilities, intentions, and willingness to cooperate to be confident that her goals will be achieved. Finally, we will look at how to synthesize abstractions that are useful for a given application/purpose.