Learning to forget : design of experiments for line-based Bayesian optimization in dynamic environments

Various scientific and engineering fields rely on measurements in 2D spaces to generate a map or locate the global optimum. Traditional design of experiments methods determine the measurement locations upfront, while a sequential approach iteratively extends the design. Typically, the cost of traveling between sample locations can be ignored, for example in simulation experiments. In those cases, the experimental design is generated using a point-based method. However, if traveling towards the next sample location incurs an additional cost, line-based sampling methods are favored. In this setting, the sampling algorithm needs to generate a route of easurement locations. A common engineering problem is locating the global optimum. In certain cases, such as fire hotspot monitoring, the location of the optimum dynamically changes. In this work, an algorithm is proposed for sequentially locating dynamic optima in a line-based setting. The algorithm is evaluated on two dynamic optimization benchmark problems.

ISBN:9781728132839

Publication year:2019

Accessibility:Closed