Doctoral research project

Person in charge of the project:
CATTHOOR FRANCKY, member of research team Associated Section of ESAT - INSYS, Integrated Systems
Proactive Hard Constraint Management in Cost-conscious Nonlinear DynamicComputing Systems
Project summary:
Modern electronic systems are increasingly dynamic with many sources of dynamism spanning multiple design abstraction levels. To minimize the system costs, dynamic systems need to be adaptive, i.e. the system should be tunable at run-time by changing some relevant system parameters, called “knobs', spanning multiple abstraction levels. One of the key design aspects of such adaptive systems is the control sub-system, which is responsible for regularly (in a demand-driven way) changing the system configuration. That should preferably happen in such a way that the target run-time cost metrics are minimized while always ensuring that the system is “reliable', i.e. it remains operational within the hard constraints imposed on it. The systems of our interest (i) are discrete with potentially nonlinear state evolution, (ii) have nonlinear and uncertainty-dependent cost functions (iii) have input-, state-, future- and uncertainty-dependent restrictions on thefeasible decision space, and (iv) have inter-dependent and mutually conflicting constraints and costs. Thus the job of the controller, is to solve at run-time a nonlinear and dynamic combinatorial optimization problem under complex constraints and uncertainties.Hiding one or many dynamics or uncertainties under worst-case abstraction is the most widely used problem simplification strategy for such systems, with the static design-time decision approach as the most extreme case. These are highly sub-optimal when the system is highly dynamic. Pure run-time approaches with simple greedy heuristics limit the controller complexity also but are known to be highly sub-optimal as nonlinearities act against them. Mixed design-time/run-time and System Scenarios are two complementary strategies from the literature to tackle the controller complexity, with each one cost-effectively handling an important class/type of dynamics.In this thesis we show that with highly varying dynamics at a fine granularity, even these advanced techniques will result in sub-optimal systems. Based on the observations that (i) nonlinearity makes look-ahead desirable to avoid timedeception and (ii) many “correlations' are present among system dynamics, we propose a truly-proactive approach to deal with such systems. We emphasize the need for Dynamic Bounding of Uncertainties and Proactive Conditioning to maximally exploit system correlations and to achieve the best possible cost minimization. We also introduce the notion of Contextual Turbo Modes, system operation modes which are only available for a short period based on the state of the system. We show how they can further optimize the cost.Based on these principles, we propose a systematic set of principles to design cost-effective controllers for the target class of systems with hard-guarantees. We illustrate the proposed reusable principles using fine-grain task scheduling in a video processing system as a case-study.
ph.D student :
Faculty of Engineering Science
Doctoral Programme in Engineering Science (Leuven)

ph.D defence : 29.06.2015
Full text ph.D