Novelty search for software improvement of a SLAM system

Genetic Improvement (GI) performs a search at the level of source code to find the best variant of a baseline system that improves non-functional properties while maintaining functionality with noticeable results in several domains. There a many aspects of this general approach that are currently being explored. In particular, this work deals to the way in which the search is guided to efficiently explore the search space of possible software versions in which GI operates. The proposal is to integrate Novelty Search (NS) within the GISMOE GI framework to improve KinectFusion, which is a vision-based Simultaneous Localization and Mapping (SLAM) system that is used for augmented reality, autonomous vehicle navigation, and many other real-world applications. This is one of a small set of works that have successfully combined NS with a GP system, and the first time that it has been used for software improvement. To achieve this, we propose a new behaviour descriptor for SLAM algorithms, based on state-of-the-art benchmarking and present results that show that NS can produce significant improvement gains in a GI setting, when considering execution time and trajectory estimation as the main performance criteria.