Efficient Evolution for Intelligent Autonomous Systems

Intelligent software is pervading the world around us. It is the driving enabler for innovations and products from autonomous vehicles, to enhanced health care, to efficient crisis response, to simple things like recommending the right movie for inspiration. However, often Artificial Intelligence (AI) does not deliver services alone but is embedded inside a larger software system.

With increasing shifts towards AI we see new challenges in the construction and evolution of software systems. AI components are often not easily adapted and changing interfaces might mean costly retraining or synthesizing of the software. This might require the augmentation or creation of new datasets and specifications.

This project will investigate how evolution can be efficiently supported for intelligent software systems. As an example, as part of the software of a self-driving car we might have an AI component that detects pedestrians, cyclists, and traffic lights. A software evolution step might change component interfaces to detecting dangers instead of the previously more detailed classification. Here the creation of new training data and costly retraining might not be necessary if we can compose the existing AI component with one that extracts dangers from classified objects.

This project will suggest suitable adaptations both on the software architecture level and on the AI level to make evolution more efficient. It is expected that this holistic view will be able to leverage techniques from software architecture and from AI that complement each other to provide timely answers to existing and emerging software evolution challenges.