<b>Bayesian Optimization for Radiation Hotspot Identification and Radiation Survey</b>

<p dir="ltr">The process of characterizing the radiation of an area, or radiation survey, is important for monitoring the danger of human occupation and estimating the deleterious effects on equipment in radioactive environments. Especially for emergencies, it is important to efficiently characterize the possible danger of an area as quickly as possible. Towards this end, robots have been traditionally utilized; as the technology matures, there is a growing desire to see these surveys to be made autonomous, using the power of machine learning (ML) to guide the characterization of the radiation field and identify hotspots more quickly. However, consistent with the consequences of misestimation, the application of ML must be principled around reporting with error and providing explainability to an operator to ensure that any ML conclusion is supported by understood evidence. Further, any ML approach must contend with the statistical difficulties of radiation fields themselves: background noise is not constant in location or space, source signal attenuates via the geometry and material make-up of obstacles, and relevant detectable signals might only sparsely fill a small area of the search domain leaving large swaths of area to be dominated by noise. We suggest the use of Bayesian Optimization (BO) as the best fit given these principles. It is a statically principled, flexible, modular, and explainable method of seeking the extrema that only exacts a reasonable computational cost in implementation. Further, we argue for extending BO aided radiation survey research, specifically for the implementation of real-world experiments. We show how augmentations to the BO process that take advantage of the framework of radiation survey can drastically increase performance but also introduce complications in explainability and computational cost. We explore how different datasets cause different biases from garden models, and we compare the performance of real-life experiments to garden model simulations. In addition, we examine various radiation survey situations and strategies with a real robot. We observe significant benefits to the highest count rate predicted, the time taken per experiment and the RMSE of the survey process, characterizing the behavior of the algorithm and we show the modularity of the BO framework for efficiently developing and implementing alternate strategies.</p>