Artificial intelligence has long been inspired by the vision of autonomous systems that not only act on the larger world, but also maintain and optimize themselves. Autonomous systems perform desired tasks to achieve desired goals continuously over a long period of time without external guidance or intervention. This requires that autonomous systems know about their true capabilities (model), reason about their course of action with respect to their current conditions (planning), and reflect on their actual behavior to determine their current conditions (diagnosis).
Autonomy can be seen as a combination of two methods: Diagnosis to determine the current condition of the system and planning a course of action to optimize system operation for the current condition.
The integration of diagnosis into regular system operation is typically realized by one of the following approaches: Alternating between explicit diagnosis and regular operation or simultaneous execution of passive diagnosis and regular operation. Alternating phases often result in long periods during which regular operation must be suspended. This is particularly true, when diagnosing complex fault scenarios, such as faults that occur intermittently. The combination of passive diagnosis with regular operation is often unsuccessful, as regular operation may not sufficiently exercise the underlying system to isolate the underlying fault.
This work introduces a new architecture, coined a Self-diagnosing Agent, which realizes the integration by a novel diagnosis paradigm called pervasive diagnosis. Pervasive diagnosis actively manipulates the course of action during operation in order to gain diagnostic information without suspending operation. Consider a system where operational goals can be achieved in multiple ways. This flexibility can be exploited to generate informative operational plans that simultaneously gather information by trading off information gain with performance objectives. Therefore active diagnosis and regular operation occur at the same time leading to higher long-run performance than an integration of regular operation with passive diagnosis or alternating between explicit diagnosis and regular operation.
The main contribution is an overall framework, which tightly integrates regular operation and active diagnosis. In particular, this contribution introduces an information criterion that defines how informative a plan is, a strategy to derive informative operational plans, and a diagnosis framework to efficiently update the belief states. The overall framework is optimized for systems with faults multiple in number, intermittent in appearance and potentially caused by hidden interactions. As a result, systems that embody the Self-diagnosing Agent architecture benefit from active diagnosis during regular operation leading to a higher, more reliable, and more robust long-run performance.
«
Artificial intelligence has long been inspired by the vision of autonomous systems that not only act on the larger world, but also maintain and optimize themselves. Autonomous systems perform desired tasks to achieve desired goals continuously over a long period of time without external guidance or intervention. This requires that autonomous systems know about their true capabilities (model), reason about their course of action with respect to their current conditions (planning), and reflect on...
»