Диссертация (1137084), страница 30
Текст из файла (страница 30)
For SM1-CM model values are as follows: − 21, − 29, − 0,0576, − 420.LM2-CM is significantly larger and consists of 63 places, 70 transitions, 154 arcs. Metrics forLM2-CM have the following values: − 72, − 2740, − 0,0175, − 16680.Figure 4.6: Sample model LM2-CMEach of these sample models is used to generate an event log with perfect fitness and valuableother characteristics. Table 4.1 shows characteristics of perfectly fitting event logs which wereprepared for these models using event log generator that has been described in Chapter 3. Whengenerating these logs we add no noise to them.Table 4.1: Perfectly fitting event logsThen, sample models have been changed in several way.
Changed models are unable to replaycorresponding generated event logs. The fitness of such a broken model and the correct eventlog is lesser than 1. Then, various repair algorithms have been applied to the changed models toevaluate them. The remainder of this chapter discusses the experiment in detail.4.3Experimental EvaluationBefore doing repair, we experimented with re-discovery. What if apply process discoveryalgorithms which guarantee perfect fitness to event logs generated from correct models? Considerthe results of this re-discovery.Two model discovered from the log 1 using Inductive and ILP miners are shown in Figure 4.7and Figure 4.8 correspondingly.A model SM1-RD-Ind is a workflow net with perfect fitness to 1 .
One may see that this modelis more or less similar to the model SM1-CM, but have 4 additional silent transitions. What is most127important, that these additional transitions present in different parts of the model. In other words,changes are not local.Figure 4.7: Model SM1-RD-Ind discovered from 1 using Inductive minerA model SM1-RD-ILP is a Petri net that fits 1 perfectly. ILP miner does not create specialoutput places. Meanwhile, a structure of the model SM1-RD-ILP looks more similar to a structureof the model SM1-CM.Figure 4.8: Model SM1-RD-ILP discovered from 1 using ILP minerIt can be concluded, that for such a small model re-discovery works relatively well. Much worseis the case, when a model is larger. In particular, Figure 4.9 shows a model LM2-RD-Ind discoveredfrom the event log 2 using Inductive miner.
This model is workflow net, as always when inductivealgorithm is used. One may see, that a structure of this model is differs from a structure of LM2-CMsignificantly.Figure 4.9: Model LM2-RD-Ind discovered from 2 using Inductive minerFigure 4.10 shows a model LM2-RD-ILP discovered from the event log 2 using ILP miner. Thisis substantially a spaghetti model that has a little in common with the model LM2-CM.
However, amodel LM2-RD-ILP fits 2 perfectly, and is able to produce all needed behaviour. Thus, re-discoveryis completely useless when models are relatively large, and a model’s structure need to be saved.The remainder describes experiments with repair algorithms. Both process models are changedin various ways. In each case, one or more pairs of transitions in the initial net are swapped. Note128Figure 4.10: Model LM2-RD-ILP discovered from 2 using ILP minerthat the modular repair technique has been proposed for repair of relatively well models.
Becauseof that, models are changed not significantly. Otherwise, when a model is completely different tothe initial one, the best choice is to completely re-discover a model using single event log. As inreal-life, patching up has its limits.Then, each changed model is repaired using all repair algorithms presented in Chapter 2.Besides, models are repaired using the method of D. Fahland and W. van der Aalst [25].
Thus, onemay compare repair technique, a re-discovery from scratch. At the end of each section, tables areshown which summarize the results of experiments.4.3.1Local Process Model RepairFirstly, we consider how local changes are repaired. In that cases, a workflow net can be repairedby replacing separate sub-nets.Model SM1-BL-1The first change is rather simple (see Figure 4.11, a change is highlighted with red colouring).An order of two neighbour transitions 4 and 6 is changed12 .
In the initial model, transition 4 isalways precedes transition 6 , whereas in a model SM1-BL-1 6 should fire before 4 . This changeis the most local, and the most simple.Figure 4.12 shows a result of repair using the naive repair technique (see Section 2.3) thatemploys maximal decomposition. In this case, results of repair using Inductive and ILP miners are12In the remainder, it is assumed that transitions have unique equivalent labels. For example, a transition 4 islabelled with the label “4 ”, whereas 6 is labelled with “6 ” etc.
Thus, we use labels to identify transitions in thetext.129Figure 4.11: Changed model SM1-BL-1the same. Note that in this figure a fragment of the model replaced by the repair is highlightedwith green. All borderline transitions of a repaired fragment are shown as greenish boxes, astransitions 4 and 6 in Figure 4.12. If a transition is inside the changed fragment, then it remainsblack-and-white, but is placed inside the green border together with inner places and arcs.Figure 4.12: Model SM1-BL-1 repaired using the naive algorithm with Inductive or ILP minersA maximal decomposition of the initial net SM1-BL-1 contains 18 sub-nets, one of which hasinconsistencies with the corresponding sub-log. This fragment contains exactly two transitions 4and 6 with a place between them.
One may see, that this is the only changed fragment. In thisand other figures, a part of model outside green border remains untouched by the repair.Figure 4.13 shows a model SM1-BL-1 repaired using the improved repair algorithm (seeSection 2.4). Note that the repaired model is equal to the correct model SM1-CM. Thus, the repairresult is as desired.One may note that the replaced fragment is larger. In particular, an unfitting sub-net of themaximal decomposition — that contains transitions 4 and 6 — is joined with two neighboursub-nets. Thus, new border transitions of the enlarged fragment are and 5 .Again, ILP and Inductive miners produce equal models.
Unfortunately, this is the case forsimplest situations only.Figure 4.13: Model SM1-BL-1 repaired using the improved algorithm with Inductive or ILPminersFitness of the event log 1 and the changed model SM1-BL-1 is 0,94, precision can be estimatedas 0,86. The naive repair technique improves fitness, but reduces precision.
Fitness between 1 andthe model from Figure 4.12 is 1, while precision is 0,57. The fragment replaced by repair contains3 nodes of 40. The model repaired using the improved algorithm have the same characteristics asthe initial (fitness to 1 is 1, precision is 0,91).
The changed fragment contains 7 nodes of 40.130For comparison purposes, we also experiment with the model repair method of D. Fahland andW. van der Aalst [25] that has been described in Section 1.3.1. The problem statement consideredby the authors allows them to apply the method in cases considered in this thesis.The method of D.
Fahland and W. van der Aalst has many settings which must be specified bya user. In particular, the following settings are implemented in the ProM 6 plug-in Model Repairthat has been used to conduct the experiments: detect loops (Y/N); detect sub-processes (Y/N);remove infrequent nodes (Y/N, keep if = 5); align sub-logs (Y/N); compute global cost function(Y/N). In this experimentations, seven combinations of settings are used which shown in Table 4.2.Table 4.2: Model Repair plug-in — combinations of settingsRepaired models obtained using all seven combinations are shown for the model SM1-BL-1.
Inthe remainder, only several models will included in text. However, all models are available at thesite of this project.Figure 4.14: Model SM1-BL-1 repaired using the Model Repair plug-in with N-N-N-N-N settingsFigure 4.14 shows a model repaired using the most basic combinations, when loops and subprocesses are not detected by the plug-in. Compare this model with a model in Figure 4.12 or amodel in Figure 4.13. It is more complex and differs a lot from the initial model SM1-CM.Figure 4.15: Model SM1-BL-1 repaired using the Model Repair plug-in with loop detectionFigure 4.15 shows a model repaired using loop detection, whereas Figure 4.16 shows a modelwith sub-processes introduced in it instead of separate transitions. These models are simpler, and131closer to the initial model SM1-CM.
Especially, it is correct for a model from Figure 4.16 that addsa single silent transitions to SM1-CM. The other side is that this model does not perfectly fits theevent log 1 . A transition 6 should fire before a transition 4 in the model, thus the model hasnot been repaired.Figure 4.16: Model SM1-BL-1 repaired using the Model Repair plug-in with sub-process detectionFigure 4.17 shows the result of repair, when the plug-in removed infrequent nodes. Actually,such a this is not the goal of this thesis, thus this setting is not used for other experiments.One may see, that the repaired model is much more complex than the initial one, and has manyunnecessary nodes.Figure 4.17: Model SM1-BL-1 repaired using the Model Repair plug-in with infrequent nodesdeletionThe result with settings Align Sub-logs and Compute Global Cost Function is shown inFigure 4.18. This model is similar to the one shown in Figure 4.17.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
