Page - 39 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Test Scenario C# Implementation ASPwithinC# Implementation Standalone ASP µ [ms] σ [ms] #TRS µ [ms] σ [ms] #TRS µ [ms] σ [ms] #TRS 1 0.00 0.00 10 415.50 18.16 10 8.30 8.92 10 2 0.30 0.48 10 2,802.20 4,445.21 10 1,428.90 2,746.91 10 3 0.00 0.00 10 / / 0 / / 0 Table 1: Runtimeandsolved test runs (TRS) for thedifferentBMWtaskassignment implementations Test Scenario C# Implementation ASPwithinC# Implementation Standalone ASP µ [ms] σ [ms] #TRS µ [ms] σ [ms] #TRS µ [ms] σ [ms] #TRS 1 0.00 0.00 10 473.80 81.24 10 13.90 8.88 10 2 16.20 4.87 10 788.10 350.75 10 341.70 404.17 10 3 1,753.30 127.58 10 / / 0 / / 0 Table 2: Runtime and solved test runs (TRS) for the different BMW park and charge assignment implementa- tions instances, regardless of the number of constraints. Though, once the size of the problem hits the com- binatorial blow-up point, it fails to return an optimal planwithin time. To measure the quality of the solution, we con- sider the two metrics we described in the previous section: most important are the overall travelling costs for task assignment and critical charging; the travelling distance of the other kinds of assignment (like for parking places) are then considered. Since the optimization strategy adopted with ASP is very similar to the one already used in the original pro- gram,inall thescenariosinwhichtheoptimaldeclar- ative solution is found within time, its quality w.r.t. to these metrics coincides to the Java solution qual- ity. For this reason, like we did for the BMW case, when the ASP solver fails to find the optimal solu- tionwithin the limit,weare interested in theanalysis of the best ASP solution found so far. This situation shows up in the third scenario. Looking at Table 4, wherethetotalcost(whichistheweightedsumofthe twometrics) is shown, it canbeseen that theoriginal implementationinJavaprovidesasignificantlyfaster and better solution than theASPimplementations. 5.Conclusion The goal of this work is to make a comparison, in different real-world logistics scenarios, between the classic imperative paradigms and the declarative ones. Answer Set Programming was chosen because of its high efficiency, as witnessed by the many ap- plications in industry. To achieve that, the FMS of BMW and IncubedIT were first analyzed, and then integrated with a new scheduler modeled in ASP. In theprevioussection, resultsandcomparisonbetween the twoapproaches inbothcompaniesareshownand analyzed. Asweexpected, there isnotaclearwinner between the two systems, but this comparison high- lighted the pros and cons of both languages, whose performance highly depend on the kind and size of tasks tobeaccomplished. Onemainqualitycriterion of the FMS is the performance and the quality of the results. To evaluate the criterion, test scenarios have been set up that are based on typical use-cases of the FMS. Regarding the BMW use-case, the imper- ative solution is significantly faster than the declara- tive one, especially for the task assignment problem. However, in ASP we make use of a different opti- mization technique, which rewards with better solu- tions . This different strategy led to a trade-off be- tween solving time and solution quality: if the im- perative method is faster, ASP finds better solutions. The IncubedIT use-case gave instead different re- sults, making clear how a very specific scenario can benefit from a particular approach rather than a gen- eral one. However, a common behavior can be seen from both BMW and IncubedIT, which represents the main weakness of ASP and enumeration tools in general. It does not scale over the size of the prob- lem. Yet in the IncubedIT scenario in which ASP doesnotexperienceacombinatorialblow-up, itfinds thebest solution in less time thanJava,withoutcom- 39