Parallel and Incremental Verification of Hybrid Automata with Ray and Verse

Abstract

Parallel and distributed computing holds a promise of scaling verification to hard multi-agent scenarios such as the ones involving autonomous interacting vehicles. Exploiting parallelism, however, typically requires handcrafting solutions using knowledge of verification algorithms, the available hardware, and the specific models. The Ray framework made parallel programming hardware agnostic for large-scale Python workloads in machine learning. Extending the recently developed Verse Python library for multi-agent hybrid systems, in this paper we show how Ray’s fork-join parallelization can help gain up to 6$\times$ speedup in multi-agent hybrid model verification. We propose a parallel algorithm that addresses the key bottleneck of computing the discrete transitions and exploits concurrent construction of reachability trees, without locks, using dynamic Ray processes. We find that the performance gains of our new reachset and simulation algorithms increase with the availability of larger number of cores and the nondeterminism in the model. In one experiment with 20 agents and 399 transitions, reachability analysis using the parallel algorithm takes 35 min on a 8 core CPU, which is a 6.28$\times$ speedup over the sequential algorithm. We also present an incremental verification algorithm that reuses previously cached computations and compare its performance.

This research was funded in part by NASA University Leadership Initiative grant (80NSSC22M0070) and a research grant from the NSF (SHF 2008883).