Диссертация (1137084), страница 35

Просмтор этого файла доступен только зарегистрированным пользователям. Но у нас супер быстрая регистрация: достаточно только электронной почты!

Текст из файла (страница 35)

In particular,its precision is 0,45. Still, both models shown in Figure 4.64 and Figure 4.65 are structurallydifferent from the model LM2-CM.Figure 4.65: Model LM2-BNL-1 repaired using the Model Repair plug-in with loop detectionEvent more structurally complex model are constructed by the Model Repair with settingscombinations 4 and 5. Note the redundant handling silent transitions in the model shown inFigure 4.66. The whole bunch of such transitions is attached to the place 5 .Figure 4.66: Model LM2-BNL-1 repaired using the Model Repair plug-in with infrequent nodesdeletionFigure 4.67 shows a model repaired using the combination 5. It also has many copied transitions16 and lots of added silent transitions. Even visually one may note that models shown inFigure 4.66 and Figure 4.67 are significantly different from the model LM2-CM.154Figure 4.67: Model LM2-BNL-1 repaired using the Model Repair plug-in with settings alignsub-logs and compute global cost functionSummaryThis section considered experiments with model repair techniques when inconsistencies arenon-local.

Characteristics of models repaired using different techniques are shown in Table 4.11.Columns of this table have the same meaning as of Table 4.8.Table 4.11: Process model repair in the non-local caseOne can easily compare the local (naive and improved) repair techniques with the non-local(greedy) method. The greedy technique constructed process models, which are as precise as themodels discovered from scratch.One of the experiments is completely unsuccessfully In particular, our implementation of thegreedy technique failed to repair the model SM1-BNL-1 using ILP miner in acceptable time.This is due to the high computational complexity of both ILP mining and alignment-basedconformance checking.

Moreover, ILP miner tends to construct spaghetti models when tryingto discover perfectly fitting one. These models are especially complex for the alignment-basedreplayer. After several iterations the fragment-to-repair became relatively large, and replay became155very time consuming due to its exponential nature. Thus, at each iteration of greedy algorithma lot of computations is performed by both mining and conformance chacking in that case. Weconclude, that this set of algorithms is not very efficient, and needs a powerful hardware.It is easy to see, that in the considered examples the greedy algorithm changes relatively largepart of the model compared with local algorithms.

For example, for the model LM2-BNL-1 thenaive approach affected 7 nodes of 133, when the greedy approach affected 106 nodes. However,the discovery of a new model completely changes the model. Besides, the greedy repair techniquehighly depends on the placement of the unfitting sub-nets in the model that needs repairing. Inthe considered examples, inconsistencies are of such type, so unfitting fragments are close to thedifferent ends of the model. The one is close to the source, whereas, the other one is close to thesink. In all such cases, a large fragment is replaced. However, these are the worst cases.Moreover, the strategy used by the greedy technique is very simple and straightforward. At eachiteration it joins all neighbour sub-nets to the fragment-to-repair.

Obviously, better strategies canbe proposed, which will reduce the number of sub-nets included in the replaced fragment. However,more complex strategies will make more operations. Thus, the repair will be more slow. We planto investigate this problem in the future, and to propose such algorithms.The modular repair technique may not add or remove labelled transitions. It only can add silenttransitions and places as a result of fragment re-discovery, if they are needed in the control-flow.That is why the repaired models have more nodes in some repair cases.Note how model repair using ILP miner is significantly faster for considered examples, than thediscovery of a completely new model even taking into account the time spent on decompositionand additional fitness checks. A repair is expectedly slower using the inductive algorithm, butinterestingly, the greedy approach is not much slower than the local repair algorithms.It is also easy to notice, that the difference between used discovery algorithms has moreinfluence in the non-local case.

The local algorithms discover only small sub-nets, thus, discoveredmodels are more or less similar.Moreover, the result of greedy repair is similar to the discovery of a new model from the wholelog. This conclusion is obvious, but has to be stated.That is why the selection of an appropriate repair algorithm for a concrete case depends onthe severity of inconsistencies.4.3.3Repair of Larger Process ModelsTo evaluate the scalability of the modular repair technique, experiments with larger workflownets have been conducted. Sample models has been made combining copies of the model LM2-CM.Figure 4.68 shows what structure these models have.

Models LMx2, LMx4, LMx8 are constructed of2, 4, and 8 LM2-CM models correspondingly. Each ellipse represents a single copy of LM2-CM model.In LMx2 and LMx4, all models are joined sequentially, whereas LMx8 has two parallel branches, eachof which is made of 4 models LM2-CM.156LMx2LM2-CMLM2-CMLMx4LM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLM2-CMLMx8Figure 4.68: Sample models to evaluate a scalability of the modular repair techniqueThe model LMx2 has 140 transitions, 125 places, and 308 arcs connecting these nodes. Model’smetrics have the following values: − 144, − 5480, − 0,0088, − 110000.The model LMx4 contains 280 transitions, 249 places, and 616 arcs connecting these nodes.Model’s metrics have the following values: − 288, − 8857, − 0.0043, − 787680.LMx8 contains 562 transitions, 500 places, and 1238 arcs connecting these nodes.

Model’smetrics have the following values: − 579, − 9524, − 0.0021, − 3150724.Figure 4.69: Model LMx2-CMFigure 4.70: Model LMx4-CMFigure 4.71: Model LMx8-CMIn this section, models are far too large to be informatively depicted within the text. Becauseof that, several figures presented to show only a big picture view. Instead of detailed figures,157characteristics of repaired model are shown in the cases when it is possible to calculate them. Acurious reader may find models themselves at the web-page of this project: http://pais.hse.ru/research/projects/iskra.For each of these model a perfectly fitting event log has been generated.

Table 4.12 showscharacteristics of these event logs which have been prepared using the event log generator describedin Chapter 3. When generating these logs, we add no noise to them. These event logs are relativelylarge in terms of containing in them different event classes, especially the event log 5 . A processwith 562 activities is complex and high-scale.

Existing software experiencing problems even whenvisualising such process models and event logs.Table 4.12: Perfectly fitting event logsThen, changed models are prepared based on fitting models. Two groups of them areprepared to repair. The first group contains model with local inconsistencies: LMx2-BL, LMx4-BL,LMx8-BL. Models with non-local inconsistencies are in the second group: LMx2-BNL, LMx4-BNL-1,LMx4-BNL-2, LMx8-BNL.Consider the results of experiments with these large models. Table 4.13 shows thecharacteristics of models constructed by the modular repair technique.Unfortunately, alignment-based conformance checking failed in most of the cases.

We haverepaired model but can not calculate their characteristics. Actually, the algorithm event fail oninitial models LMx4-CM and LMx8-CM. Thus, we show the following characteristics: size of themodel (a number of nodes); an overall number of sub-nets in models’ decomposition (N-f); anumber of sub-nets with inconsistencies (N-bf); a number of nodes of all replaced sub-nets (a sizeof the fragment-to-repair, Ch.). These characteristics show how large the models is, and how themodel repair influences on a repaired model.To conduct experiments with larger models, we used the following more productive hardwareconfiguration: central processing unit Intel Core i7-7700 with a clock rate of 3.60 GHz ; 16 Gbof DDR4 random access memory (2400 MHz ); Windows 10 x64 operating system; and combinedSSD/HDD drive.

Характеристики

Список файлов диссертации