Automated generation of consistent models using qualitative abstractions and exploration strategies

Area of Science:

• Computer Science
• Artificial Intelligence
• Software Engineering

Background:

• Automated model synthesis is vital for testing autonomous driving systems.
• Existing solvers struggle with scalable, simultaneous handling of structural and attribute constraints, especially geometric ones.
• Inconsistent models lead to irrelevant test cases (e.g., false positives).

Purpose of the Study:

• To develop a scalable method for automatically synthesizing consistent models for autonomous driving testing.
• To ensure synthetic models satisfy complex structural, attribute, and geometric constraints.
• To maintain theoretical properties like completeness and diversity during model generation.

Main Methods:

• Integration of a structural graph solver (using partial models) with an SMT-solver and a quadratic solver.
• Implementation of sophisticated bidirectional interaction between solvers for consistency checks, decision, and propagation.
• Introduction of custom exploration strategies to accelerate model generation.

Main Results:

• The proposed approach successfully derives models that simultaneously satisfy structural and numeric constraints.
• The method ensures key theoretical properties like completeness and diversity are maintained.
• Evaluation demonstrates scalability and diversity across four complex case studies.

Conclusions:

• The combined solver approach offers a scalable solution for synthesizing consistent models in autonomous driving.
• This method enhances the reliability and coverage of test cases by addressing complex constraints.
• Custom exploration strategies further optimize the model generation process.