Scenic: a language for scenario specification and data generation

Abstract We propose a new probabilistic programming language for the design and analysis of cyber-physical systems, especially those based on machine learning. consider several problems arising in process, including training system to be robust rare events, testing its performance under different conditions, debugging failures. show how can help address these by specifying distributions encoding interesting types inputs, then sampling generate specialized test data. More generally, such languages used write environment models, an essential prerequisite any formal analysis. In this paper, we focus systems as autonomous cars robots, whose at point time is scene , configuration physical objects agents. domain-specific language, Scenic describing scenarios that are over scenes behaviors their agents time. combines concise, readable syntax spatiotemporal relationships with ability declaratively impose hard soft constraints scenario. develop techniques from resulting distribution, taking advantage structure provided ’s syntax. Finally, apply multiple case studies training, testing, neural networks perception both standalone components within context full system.