Surrogate-Based Decision Boundary Evolutionary Sampling Method for Autonomous Systems

Searching the decision boundary of an autonomous system is essential for understanding its behavioral characteristics. In this paper, the decision boundary search problem is mathematically modeled as an optimization problem with continuous inputs and discrete outputs, which aims at searching continuous regions. A surrogate-based decision boundary evolutionary sampling method for autonomous systems (SDBES) is proposed. SDBES combines the evolutionary algorithms with adaptive sampling methods, enabling the samples to adaptively search and cover decision boundaries in iterative evolution. The subject of evolutionary iteration shifts from individual to community. The adaptive grouping strategy is designed to increase the diversity of sample points. The elite-pool strategy and elimination strategy are used to improve convergence. To fully compare the performance of algorithms with uniform standards and low cost, we design a series of test cases and two new metrics that consider repeatedly covered true boundary points and the number of queries of actual systems. The ablation experiments are conducted to verify the effectiveness of the strategy. SDBES is compared with two state-of-the-art adaptive sampling algorithms under the proposed 17 types of benchmarks and various sample numbers. Experimental results show that the sample points generated by SDBES could be concentrated in decision boundary regions with high precision and efficiency. The test of eight path planning scenarios illustrates that SDBES can effectively find the decision boundary of real-world autonomous systems and provide a new method for their robustness test.