SEA4DQ Workshop 2022

A self-adaptive system can modify its structure and behavior at runtime based on its perception of the environment, itself, and its requirements. By adapting itself at runtime, the system can maintain its requirements in the presence of dynamic environment changes. Examples are elastic cloud systems, intelligent IoT systems as well as proactive process management systems. One key element of a self-adaptive system is its self-adaptation logic, which encodes when and how the system should adapt itself. When developing the adaptation logic, developers face the challenge of design time uncertainty. This means they have to anticipate potential environment states and the precise effect of adaptation in a given environment state, while the knowledge available at design time may not be sufficient to do so. A recent industrial survey determined design-time uncertainty as one of the most frequently observed difficulties in designing self-adaptation logic in practice. This talk will explore the opportunities but also challenges that modern machine learning algorithms offer in building the self-adaptation logic in the presence of design-time uncertainty. It will focus on online reinforcement learning as an emerging approach, which means that during operation the system learns from interactions with its environment, thereby effectively leveraging data only available at run time. In particular, the talk will focus on three different issues related to data quality and will introduce initial solutions for these issues: (1) data non-stationarity, (2) data sparsity, and (3) data intransparency. The talk will close with a critical discussion of limitations and an outlook on future research opportunities.

Nowadays, we are witnessing an increasing demand in both industry and academia for exploiting Deep Learning (DL) to solve complex real-world problems. However, the performance of these high-capacity learners is currently bounded by the quality and volume of their underlying training data. The use of incomplete, erroneous, or inappropriate training data, and the implementation of poor data management practices in a training pipeline often result into unreliable, biased, or under specified models. In this talk, I will report about some recent research works that we have conducted to identify best practices of data management for DL. I will also report about recent techniques and tools that we have developed to help detect the root cause of model under-specification issues early on during a DL training process.