CS MSc Thesis Presentation 27 November 2023
One Computer Science MSc thesis to be presented on 27 November
Monday, 27 November there will be a master thesis presentation in Computer Science at Lund University, Faculty of Engineering.
The presentation will take place in E:2405 (Glasburen). Please see link below.
Note to potential opponents: (Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@example.org). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.)
10:15-11:00 in E:2405 (Glasburen)
Presenter: Oskar Andersson
Title: Unsupervised Learning-Based Test Scenario Selection using Autonomous Vehicle Disengagements
Examiner: Emelie Engström
Supervisor: Qunying Song (LTH)
Autonomous vehicles are becoming a subject of testing on public roads to ensure their safety before making them publicly available. With the complexity of their operation, new testing routines and standards needs to be implemented and evaluated to ensure safe operation. Many research papers on the subject apply a scenario-based testing methodology with the principle of finding representative test scenarios and ensuring that the system performs appropriately for these. Currently the field is faced with challenges in articulating test scenarios and making sure that they capture all possible scenarios. This thesis use a scenario-based testing approach combined with unsupervised learning to find representative scenarios automatically from realistic autonomous vehicle disengagements. The resulting clusters are evaluated to determine what form of vectorization and embedding of textual entries leads to the most accurate results. The results from clustering were that the methodology was able to produce clusters with high performance in regards to three common clustering metrics for a data set of 184 disengagement entries. The evaluation of the actual scenarios that the methodology was able to cover did however not indicate that the methodology achieved a high level of accuracy, with the highest percentage achieved being approximately 41% for KMeans clustering with 19 clusters and approximately 53% coverage with 35 clusters using DBScan. In conclusion, the report reveals the methodology as a feasible way to mine test scenarios. However, the lack of large data sets of disengagements makes the tool hard to conclusively evaluate and the similarity comparison between disengagement scenarios is hindered by the lack of embeddings specialized in semantics in the field of autonomous vehicles.
Link to popular science summary: To be uploaded
From: 2023-11-27 10:15 to 11:00
birger [dot] swahn [at] cs [dot] lth [dot] se