Tags:benchmarking, connection method, description logic, optimizations and reasoning
Abstract:
DL reasoners were developed with cutting-edge performance, implementing plenty of specific optimization techniques over tableaux based methods, which took over the field for years. On the other hand, promising methods may have been neglected in such a scenario, in which the tough competition is often focused on gains through optimizations. Therefore, perhaps there is still room available for “basic research” on DL reasoning, in the sense that other efficient calculi need to be adapted to DL, tuned and tested. The purpose of this work is to stimulate research on trying out DL calculi (other than tableaux) by making a careful, detailed comparison between tableaux and other inference methods in a systematic way: first starting with simpler languages (like ALC) without any optimizations. Then gradually including optimizations and comparing them; and continuing these interactive steps: enhancing language expressivity, including optimizations, and testing until reaching the last advances on optimizations and the most expressive DL fragments such as SROIQ. The comparison can also be done by in terms of two other additional aspects: memory usage and algorithm asymptotic analysis, with worst and average cases, etc. The rationale is identifying whether there are fragments which are more suitable to certain inference methods, as well as which aspects or constructs (e.g., the costliest combinations, which usually involve inverses, nominals, equalities, etc) are sensitive to which calculus. Concretely, the idea is to start out with our DL connection calculi [2,3] and reasoner, the RACCOON (ReAsoner based on the Connection Calculus Over ONtologies) [4], which for now takes on ALC, the basic DL fragment, and displayed a surprisingly effective performance in comparison with other well-established reasoners, such as Hermit, FaCT++ and Konclude.
A Roadmap to Gradually Compare and Benchmarking Description Logic Calculi