Bayesian Forensic DNA Mixture Deconvolution With a Novel String Similarity Measure

Forensic scientists have great interest in deconvolving profiles in crime scene samples, particularly to test the presence of a courtroom defendant. However, DNA degradation, low material quantity, and mixed signals from multiple contributors pose constant challenges. Additionally, crime scene analysis includes DNA amplification which generates artifact sequences, confounding the data. Capillary electrophoresis is used worldwide, but massively parallel sequencing technology enables data analysis at a much higher resolution, increasing discrimination between profiles in the sample. This modern data type requires new statistical methods to ensure rigorous analysis. In this work we propose a Bayes factor that tests whether a defendant's DNA was present at a crime scene. The model is able to include known contributors, such as the victim. As part of this methodology, a novel string edit distance was published to measure similarity between alleles and their sequencing artifacts. This edit distance accommodates stutter, a common artifact seen after amplifying forensic loci. The Bayes factors produced by the model demonstrate the ability to detect the suspect, with variability in the estimate depending on the defendant's mixing proportions. When the person of interest is not present, there is decisive rejection with no false positives. In another area of machine learning, off-road autonomous vehicles face a unique set of challenges compared to their road-going counterparts. Lane markings and road signs are unavailable, with mud, vegetation, and steep slopes taking their place. Particularly, autonomy struggles determining whether vegetation or what it obscures is drivable. Modeling and simulation of sensors enhances and accelerates autonomy development, but analytical models in literature and bespoke simulations did not agree on how well lidar penetrates grass-like vegetation. To test our simulator against the analytical model, we constructed vegetation mock-ups that conformed to the assumptions of the analytical model and measured pass-through rate on calibrated lidar targets. Distance, vegetation density, and target reflectivity were varied. A random effects model was used to address the dependence introduced by repeated measures, which increased accuracy while reducing time and cost. Results varied by lidar unit, with significant three-way interaction. Results suggest benchmarking experiments could be useful in autonomy development.