Incremental Learning Through Graceful Degradations in Autonomous Systems

Principal Investigator: Bharat Bhargava

Intelligent Autonomous Systems (IAS) are highly cognitive, reflexive, multitasking, trustworthy (secure as well as ethical), and rich in knowledge discovery. IAS are deployed in dynamic environments and connected with numerous devices of different types, and receive large sets of diverse data. They often receive new types of raw data that was not present in either training or testing data sets thus they are unknown to the learning models. In a dynamic environment, these unknown data objects cannot be ignored as anomalies. Hence the learning models should provide incremental guarantees to IAS for learning and adapting in the presence of unknown data. The model should support progressive enhancements when the environment behaves as expected or graceful degradations when it does not. In the case of graceful degradations, there are two alternatives: (1) weaken the acceptance test of data object (operating at a lower capacity) or (2) replace primary system with a replica or an alternate system that can pass the acceptance test. In this paper, we provide a combinatorial design— MACROF configuration—built with balanced incomplete block design to support graceful degradations in IAS and aid them to adapt in dynamic environments. The architecture provides stable and robust degradations in unpredictable operating environments with limited number of replicas. Since the replicas receive frequent updates from primary systems, they can take over primary system’s functionality immediately after an adverse event. We also propose a Bayesian learning model to dynamically change the frequency of updates. Our experimental results show that MACROF configuration provides an efficient replication scheme to support graceful degradations in autonomous systems.


Other PIs: Jason Kobes at NGC

Students: Ganapathy Mani, Basvesh Shivakumar

Representative Publications

  • submitted to IEEE Cloud 2018

Keywords: congnitive, graceful degradation, knowledge discovery, learning models, reflexive