Диссертация (1137084), страница 34
Текст из файла (страница 34)
The improved repairtechnique using Inductive miner produces an unsound model (see Figure 4.51) with manyadditional silent transitions. This model is not of practical value.Figure 4.51: Model SM1-BNL-1 repaired using the improved algorithm with Inductive miner148The model repaired using the ILP miner is shown in Figure 4.52 This model is more similar tothe initial model SM1-CM and is able to replay all traces of the log 1 .
Nevertheless, it has the samedisadvantages as the model shown in Figure 4.50. In particular, a transition 20 has no outgoingarcs, whereas 0 has no incoming arcs. Because of that 0 can e fired at a needed step of the replay,but also at other steps as well.Figure 4.52: Model SM1-BNL-1 repaired using the improved algorithm with ILP minerThe greedy technique to repair non-local inconsistencies of that type has been describedin Section 2.5.
Figure 4.53 shows the model SM1-BNL-1 repaired using this greedy technique.Actually, the algorithm made several iterations and included all fragment of decomposition intothe fragment-to-repair. In that case, a model repair reduced to re-discovery. The repaired modelis the same as has been shown in Figure 4.7. The characteristics of these models are equal. Thus,a model SM1-BNL-1 is one of the models which are most complex to be repaired by the greedytechnique. One may use ILP miner to obtain a model equal to the one shown in Figure 4.8.Figure 4.53: Model SM1-BNL-1 repaired using the greedy technique with Inductive minerUnlike in local cases, in non-local cases of such a type repairs of the ProM 6 Model Repair plugin look better in comparison with the results of naive and improved repair techniques presented inthis thesis.
Table 4.9 shows the characteristics of the repairs of the model SM1-BNL-1 using sevensettings configurations. Configurations 1, 2, 4, and 5 repaired the model perfectly. Consider theseresults.Table 4.9: Characteristics of SM1-BNL-1 repairs using the Model Repair plug-in149Figure 4.54 shows how the model SM1-BNL-1 is repaired using the settings combination 1. Inthis model, transitions 0 and 20 are replaced as it is needed. Besides, silent transitions are addedto the model. One of them is totally redundant.Figure 4.54: Model SM1-BNL-1 repaired using the Model Repair plug-in with N-N-N-N-N settingsThe result of repair using the combination 2 is shown in Figure 4.55.
This model is eventsmaller than the initial model SM1-CM. At the same time, fitness between this model and the eventlog 1 is 1. However, it is less precise than SM1-CM.Figure 4.55: Model SM1-BNL-1 repaired using the Model Repair plug-in with loop detectionRepairs using settings combinations 4 and 5 are more complex. Figure 4.56 the model repairedusing the ProM 6 Model Repair plug-in with infrequent nodes deletion. This model has a lot ofunnecessary nodes, including handling silent transitions. The model is event less precise than theone repaired using the settings combination 2.
Thus, combination 4 is not the best choice.Figure 4.56: Model SM1-BNL-1 repaired using the Model Repair plug-in with infrequent nodesdeletionThe model repaired using the settings combinations 5 (see Figure 4.57) is similar to the onerepaired using the combination 1.
Both models 4 and 5 have similar characteristics. Other settingscombination (3, 6, and 7) can not repair model SM1-BNL-1 to perfect fitness with the event log1 , and achieve only fitness of 0,88. Thus, we do not show these models here.The result obtained using loop detection (see Figure 4.55) is the most suitable for practicecases. Note how this model is similar to the re-discovered model (see Figure 4.53). Thus, it canbe concluded that for such non-local cases best repair technique is not yet proposed.
All existingmethods have disadvantages. When repairing relatively small models, their best results are similarto a result of complete re-discovery.150Figure 4.57: Model SM1-BNL-1 repaired using the Model Repair plug-in with settings alignsub-logs and compute global cost functionModel LM2-BNL-1Consider non-local repair experiments with larger models. Figure 4.58 shows the larger changedmodel LM2-BNL-1. Again, the introduced inconsistency is relatively insufficient, but non-local.
Inparticular, labels of transitions 16 and 05 are swapped. A fitness between the initial modelLM2-BNL-1 and the event log 2 is 0,98, whereas precision can be estimated as 0,59. Thus, thechanged model is as precise as the correct model LM2-CM.Figure 4.58: Changed model LM2-BNL-1As with the model SM1-BNL-1, the naive technique repairs model through replacing transitionsat the borderline of unfitting fragments. These transitions remain without incoming or outgoingarcs because a fragment does not include the whole repaired inconsistency.
For example,Figure 4.59 shows the model repaired using the naive technique with ILP miner.Figure 4.59: Model LM2-BNL-1 repaired using the naive algorithm with ILP minerIt can be easily seen, that the repaired fragments — which are highlighted with green colouras earlier — are disintegrated into separate transitions. Note how the repair constructed disjoinPetri net, in which transitions 16 and 05 causing the inconsistency are totally separate from theother model.
These transitions can fire at any step of model’s execution, they have no constraints.Because of that, the repaired model is imprecise and can be hardly useful in practical cases. Inparticular, its precision is 0,19, whereas precision of the initial model LM2-CM is 0,59. Results ofthe naive repair using inductive algorithm are event more spaghetti-like. Thus, we may conclude,that the naive repair technique is totally useless when repairing non-local inconsistenciesConsider how the improved technique repaired the model LM2-BNL-1.
Figure 4.60 shows themodel constructed by this technique with Inductive miner. The result of the improved techniquewith ILP miner is illustrated in Figure 4.61.151Figure 4.60: Model LM2-BNL-1 repaired using the improved algorithm with Inductive minerAgain, touched by the repair procedure fragments are highlighted with green colour, borderlinetransitions are highlighted with green frames. The improved technique better preserved structure ofboth models. Besides, they are more precise (0,32) than models repaired using the naive technique.Still, the inconsistency is not fully covered by a replaced fragment. Because of that, both modelscontain transitions without incoming and outgoing arcs.
The model repaired using ILP miner (seeFigure 4.61) have one disjoint transition 16 . Thus, the improved is insufficient for that non-localcase.Figure 4.61: Model LM2-BNL-1 repaired using the improved algorithm with ILP minerConsider how the greedy technique repairs this model. Figure 4.62 shows the model LM2-BNL-1repaired using this technique with Inductive miner. The changed fragment is highlighted with greenframe. Note how this model is similar to the re-discovered model SM1-RD-Ind shown in Figure 4.7.The greedy approach made iterations until not included all unfitting fragment into the single subnet. Then, this sub-net is re-discovered. The preserved fragment is relatively small and contains27 nodes.Figure 4.62: Model LM2-BNL-1 repaired using the greedy technique with Inductive minerEven more spaghetti-type model is constructed by the greedy technique using ILP miner.
Thismodel is shown in Figure 4.63. The re-discovered fragment contains 106 nodes of total 133 nodesin the model LM2-BNL-1. This model is also very similar to the completely rediscovered by ILP152miner (see Figure 4.8). Unfortunately, the model is not very structurally similar with the initialmodel LM2-CM.Figure 4.63: Model LM2-BNL-1 repaired using the greedy technique with ILP minerNote that the case is considered when unfitting fragments are close to source and sink places ofthe initial workflow net LM2-BNL-1. As we showed, in that case the greedy technique is squeezedto include almost the whole model in the fragment-to-repair.
Obviously, this is not always thecase. For example, in the case shown in Figure 2.18 approximately one half of the model preserveduntouched by the repair. Thus, even the greedy technique is the more suitable, the more localinconsistencies are.Table 4.10: Characteristics of LM2-BNL-1 repairs using the Model Repair plug-inIn that case, the ProM 6 Model Repair plug-in also produce complex models with lots ofelements, but in comparison with the results of the modular repair its position are much stronger153than in previously considered cases.
Characteristics of models repaired using the Model Repairplug-in are shown in Table 4.10. Four settings combinations perfectly repaired models: 1, 2, 4,and 5.Figure 4.64 shows the model repaired by the Model Repair plug-in. As all repairs in this case,this model contains many additional silent transitions. There are 13 copies of the transition 16in different parts of the model. All these transitions may be skipped through silent transitions, orthey are handling loop transitions.Figure 4.64: Model LM2-BNL-1 repaired using the Model Repair plug-in with N-N-N-N-N settingsThe model repaired using the Model Repair plug-in with loop detection is shown in Figure 4.65.This model perfectly fits the event log 2 , but it is less precise than the model LM2-CM.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
