Performance heuristics for GR(1) synthesis and related algorithms

Abstract

Reactive synthesis for the GR(1) fragment of LTL has been implemented and studied in many works. In this work we present and evaluate a list of heuristics to potentially reduce running times for GR(1) synthesis and related algorithms. The list includes several heuristics for controlled predecessor computation and BDDs, early detection of fixed-points and unrealizability, fixed-point recycling, and several heuristics for unrealizable core computations. We have implemented the heuristics and integrated them in our synthesis environment Spectra Tools, a set of tools for writing specifications and running synthesis and related analyses. We evaluate the presented heuristics on SYNTECH15, a total of 78 specifications of 6 autonomous Lego robots, on SYNTECH17, a total of 149 specifications of 5 autonomous Lego robots, all written by 3rd year undergraduate computer science students in two project classes we have taught, as well as on benchmarks from the literature. The evaluation investigates not only the potential of the suggested heuristics to improve computation times, but also the difference between existing benchmarks and the robot’s specifications in terms of the effectiveness of the heuristics. Our evaluation shows positive results for the application of all the heuristics together, which get more significant for specifications with slower original running times. It also shows differences in effectiveness when applied to different sets of specifications. Furthermore, a comparison between Spectra, with all the presented heuristics, and two existing tools, RATSY and Slugs, over two well-known benchmarks, shows that Spectra outperforms both on most of the specifications; the larger the specification, the faster Spectra becomes relative to the two other tools.

Spectra: a specification language for reactive systems

“…Finally, the evaluation in [28] also shows that thanks to these heuristics, Spectra outperforms two previously published tools, RATSY [6] and Slugs [22], on large specifications. While for small specifications, the other tools running times for realizability checking are better than Spectra's, the larger the specification, the faster Spectra becomes relative to the two other tools.…”

“…As empirically shown in [28], the combination of all these heuristics provides improvements in realizability checking running times. Moreover, the greatest improvement is for specifications that have slower original running times.…”

“…The algorithm for realizability checking follows [10], but, importantly, incorporates a number of performance heuristics from [28]. We have implemented heuristics in two levels.…”

Inherent vacuity for GR(1) specifications

“…We have implemented detection for the different types of vacuity as an extension of Spectra [33,42], based on CUDD [41] as a BDD library. Our implementation includes also the computation of vacuity core, as an instance of the generic DDMin algorithm implemented in Spectra (with the performance heuristics described in [19]).…”

“…We use the benchmark SYNTECH15 [19], which includes a total of 78 specifications of 6 autonomous Lego robots, written by 3rd year undergraduate computer science students in a project class taught by the authors of [19]. Out of the 78 specifications in SYNTECH15, we use 14 unrealizable ones, which we label SYN15U, and all 61 realizable ones, which we label SYN15R.…”