Dynamic Architectures for Computational Self-Awareness

近年来，人们对探索技术系统及其如何表达自主行为作为应对各种挑战的一种方式越来越感兴趣；现代系统正被用于日益多样化、不可预测和动态的环境中，这使得系统需要通过推理和学习来克服未知的状态和情况--而不需要人为干预。由于系统的环境和目标更多地受时间变化的影响，程序员在设计系统本身时不可能知道解决运行时问题的“正确”方法；因此，系统具有以有效和高效的方式自主适应响应这些变化的能力变得越来越重要。解决这些问题的一种方法是设计具有计算自我意识能力的系统，这样他们就可以在没有任何先验知识的情况下学习和行动，同时随着时间的推移扩展他们的技能。因此，这个项目旨在研究如何将自我意识计算机系统设计为在不断变化的环境中运行，特别是关注系统之间交互的额外复杂性，以及如何将这些设计为高效和有意义的。此外，重点是系统与系统之间的相互作用，而不是更大范围的社会技术系统，在那里系统和人类之间存在着复杂的相互作用。这个项目旨在解决的一个问题是，系统如何能够在自己的行为受到他人影响的环境中学习和实现它们的未来目标--无论是隐含的还是明确的。研究人员提出了关于集成问题的问题，在系统本身的设计中考虑了明确的组织；然而，这个项目的目的是调查两个交互系统之间可能出现的意外交互，以及如何设计系统来自主地解决这些问题。这涉及到考虑一个人如何既塑造自己又塑造他人，以及这些所谓的“社会制度”如何能够更深入地了解它们在周围世界中的位置，以及其他人可能对自己产生的影响。人类心理学和社会学的理论将被用来作为回答这些问题的灵感。由此产生的另一个问题涉及系统可以展示的学习过程本身，以及这些过程如何与系统运行的背景相关联。例如，某一行为的结果可能会根据背景/情况而变化；如果系统还不够复杂，不足以感知这种变化的原因(即，另一个人的行动、环境的变化)，那么系统就会认为没有理由进行这种变化，从而导致系统相应地改变其知识。这样做的一个后果是，由于未知的相互作用，可能会改变或丢失所学到的、成功的知识和行为。因此，该项目还旨在调查系统的学习过程是否可以模块化，以便背景变化不会干扰学习的知识。通过这种方式，由于采用了更健壮的学习方法，系统可能能够更有效地交互。

There has been growing interest in recent years regarding the exploration of technical systems andhow they can express autonomous behaviour as a way of addressing a wide range of challenges;modern systems are being used in increasingly diverse, unpredictable and dynamic environments,creating a need for systems to overcome unknown states and situations through reasoning andlearning - without human intervention. As the environments and the goals of the systems are morecommonly subject to change over time, it is becoming impossible for programmers to know the"correct" approach to solving run-time problems whilst designing the systems themselves; it istherefore increasingly important for systems to have the ability to autonomously adapt to respondto these changes in an effective and efficient manner. One such approach to tackling these issues is designing systems with computational self-awareness capabilities, so they can learn and act withoutany prior knowledge, whilst extending their skills over time.With this in mind, this project therefore aims to look at how self-aware computing systems can bedesigned to operate in an ever-changing environment, specifically looking at the additionalcomplexity of interactions between systems and how these can be engineered to be efficient andmeaningful. Further, the focus is on system to system interactions, rather than the larger area ofsociotechnical systems where there are complex interactions between systems and humans alike.One question this project aims to address surrounds how systems are able to learn and achieve theirown goals in an environment where their own actions are affected by another - either implicitly or explicitly. Researchers have asked questions about issues of integration which considers explicitissues in the design of the system itself; this project however aims to investigate the unintendedinteractions that may arise between two interacting systems, and how systems can be designed tocombat these issues autonomously. This involves consideration of how one models both itself andothers, and how these so-called "social systems" are able to develop a deeper understanding of their place in the world around them, as well as the impact others can have on themselves. Theories of human psychology and sociology will be used as inspiration to answer these questions.A further question that arises from this regards the learning processes themselves that a system can exhibit, and how these relate to the context systems operate in. For example, the outcome of a certain behaviour can change depending on the context/situation; if the system is not complex enough to perceive the cause of this change (i.e. the actions of another, a change in the environment), it will appear to the system that there is no reason for this change, resulting in the system changing its knowledge accordingly. A consequence of this is that well-learnt and successful knowledge and behaviour may be changed or lost due to unknown interactions. This project therefore also aims to investigate whether the learning processes of systems can be modularised, such that contextual change does not interfere with learnt knowledge. This way, systems may be able to interact more effectively due to a more robust learning approach.