Online Experiment-Driven Learning and Adaptation

This chapter presents an approach for the online optimization of collaborative embedded systems (CESs) and collaborative system groups (CSGs). Such systems have to adapt and optimize their behavior at runtime to increase their utilities and respond to runtime situations. We propose to model such systems as black boxes of their essential input parameters and outputs, and search efficiently in the space of input parameters for values that optimize (maximize or minimize) the system’s outputs. Our optimization approach consists of three phases and combines online (Bayesian) optimization with statistical guarantees stemming from the use of statistical methods such as factorial ANOVA, binomial testing, and t-tests in different phases. We have applied our approach in a smart cars testbed with the goal of optimizing the routing of cars by tuning the configuration of their parametric router at runtime.