Диссертация (1137084), страница 11
Текст из файла (страница 11)
A triple (, , ) is calculated for each model, in which is the structural similaritybetween models, stands for behavioural similarity (models are behaviourally similar if they canreplay each other’s behaviour), and stands for so-called model badness w.r.t. a given set ofinitial model runs. If the model violates one of the three soundness conditions for at least one runfrom , indicates that.A sequence of basic change operations is performed during error correction. Arcs, places, and -transitions can be added or removed.
Labelled transitions are not involved in the correctionprocedure.Then the set of corrected sound models are built, which are constructed from initial modelusing minimal number of changes. The authors use the method of simulated annealing to find theoptimal set of repair operations. An averaged triple (,,) for all candidate models is used asan objective function. All corrected models should be at least as sound as the initial model, andmore or less structurally similar to it. Note that the value of an objective function is calculatedon a random sub-set of all possible runs of the model being corrected. An information from eventlogs is not used here.1.3.6Other Model Repair TechniquesOther model repair problem statements and corresponding techniques have also beenconsidered in literature.
We will not discuss them in detail. We will only give a brief overview sincethese statements are significantly different from our problem statement, or repeat the statementswhich already discussed in this section.F. Basile, P. Chiacchio, and J. Coppola [116] proposed how to repair temporal anomalies inprocess models. Given a Time Petri net, modelling a discrete system, and a recorded system’sbehaviour (a timed sequence of events), the goal is to tune lower and upper time bounds of46specified transitions, in which they can fire. The authors apply integer linear programming tosolve the problem.
Note that this technique does not change a structure of the initial model.At the same conference (IFAC 2015), U. Martínez-Araiza and E. López-Mellado [117] havepresented their approach for CTL10 model repair. The authors have considered bounded anddeadlock-free labelled Petri nets. These are system’s models to repair. Given a system specification(a CTL formula) and a system’s model (a Petri net), the goal is to change the model in such away that it satisfies the specification. In their paper, the set of possible repair operations on aPetri net contains add and remove operations on the sets of transitions, places, and arcs (flowrelation).
In other words, basic change operations for graphs are used.A. Awad, G. Decker, and N. Lohmann [119] have considered the problem of data-aware modelsrepair in 2009. In their paper, the BPMN models with data objects are repaired. These modelsare formalised using Petri nets. The proposed approach identify modelling mistakes leading todeadlocks. For example, data preconditions of different tasks can be in coflict. The authors haveshown how these problems can be identified and repaired using their automated technique.In 2016, D. Sánchez-Charles, M. Solé, J.
Carmona, and V. Muntés-Mulero [120] have proposeda method for improvement of process model precision. They have developed a technique, whichhelps to improve the precision of iterative process models w.r.t an event log. The method is basedon the partial unrolling of loops in the model. In other words, if some behaviour tends to berepeated twice, it is more precise to have it sequentially repeated twice in the model rather thanas a loop. This technique is useful, if applied carefully.
It can be considered to be a model repair,improvement, or simplification method.Y. Sun, Y. Du, and M. Li [121] have proposed another method for repair of workflow netsw.r.t to event logs. For any workflow net a special matrix can be built. It shows the orderingrelations between process activities encoded with labelled transitions of the model. In particular,authors consider the following six relations: direct following (>), indirect following (>>), directcausality (→), indirect causality (⇒), mutual exclusion (♯), concurrency (∥).
A similar matrix isconstructed based on the event log, where activities are related to events. Y. Sun et al. call thesematrices mirroring matrices.The authors compare these matrices for a workflow net and en event log to find deviationsbetween the process and the corresponding model. The approach is of use when we need tointroduce new activities, present in the log, into a process model. Additional transitions canbe inserted between existing transition or as a new process branch. The model, repaired usingthis approach, looks very similar to the model repaired using the technique of D. Fahland andW.
van der Aalst [25].Finally, two techniques of process model repair have been proposed recently, in 2018. X. Zhanget al. [122] have developed a method to construct what they call logic Petri nets. A transitionof such a Petri net may be assigned two “logic functions”: input and output. The firing rule ismodified for these logic transitions. A transition can fire only if its input function is true. The10Computation tree logic which has been proposed by E. Clarke and E. Emerson [118].47authors propose to repair a Petri net using an event log information by assigning to a net transitionlogic functions which will allow to replay the log. Thus, as a result of this repair, an initial Petrinet is replaced by a significantly more sophisticated model.Yet another approach to repair workflow nets has been proposed by H.
Qi, Y. Du, L. Qi, andL. Wang [123]. The authors have considered the special case when new choices are introduced intothe process. These new choices leave footprints in the event log used to repair a process model.A special type of “extended” alignments is proposed in the paper that allows to find choices.The authors have shown that the results of this search can be used to build in a model severaltypes of choice structures precisely. The problem statement is rather specific and narrow.
Theirpaper can be considered as an elaboration of the model repair technique of D. Fahland andW. van der Aalst [25].1.3.7Process Model SimplificationThe problem of the automatic process model simplification is closely related to model repair.The goal is to change a process model such that it still fits the important (specified) part ofa behaviour from an event log, but contains less modelling elements.
Ideally, the simple modelcontains only sufficient modelling elements and nothing else.The most obvious simplification is to remove infrequently used parts of the model [124]. Thisoperation simplifies model, although (hopefully, insignificantly) reduces the conformance to theevent log. However, significantly more refined techniques have been presented by D. Fahland andW. van der Aalst [124] in 2013, and by J. de San Pedro, J. Carmona and J.
Cortadella [125] in2015. Earlier, in 2007 C. Gunther [44] has considered a method for simplification of fuzzy processmodels by removing the non-essential parts of the model using frequency metrics.The model simplification techniques may be used after the model repair, when a perfect fitnessis ensured but a model is over complicated.1.4ConclusionsIn this chapter we have discussed the background of our work. The field of process mining hasbeen shortly described. Basic definitions of a workflow net, an event log, and other entities havebeen presented.
Then, we have discussed recent developments in process discovery and conformancechecking, which are employed in our repair approach.Section 1.3 have considered various statements of process model repair problem with theirspecial aspects. It is easy to see, that this problem is vital and relevant. Many techniques havebeen proposed to process models in different notations and to deal with various errors in thesemodels.
The problem statement that takes the event log data has been firstly proposed in 2011 [6].Later, many other ways to reconstruct a model such that it will conform the log have beensuggested. However, most of them deal with cases when new activities are added to the process,48and corresponding transitions are to be built in the model. The other popular case is when a modelbecomes obsolete and contains transitions preventing from replaying the event log , and thereforethese transitions are to be removed from the model.
When dealing with these cases, activities arereplaced within a process, but not added or removed, the considered techniques do not always meetexpectations. In this thesis, we aim to close this issue, and propose a technique suitable for thesecases.The next Chapter 2 begins with the strict statement of the problem considered in this thesis,which differs from the statements discussed in this chapter. Then, the next chapter presents theapproach for process model repair that uses information from an event log and is based on modeldecomposition. Necessary formal proofs and techniques features are explained in detail.
Then, theremainder of the thesis considers an experimental evaluation of the presented technique. Chapter 3describes a method for artificial event log generation used to evaluate our process model repairexperimentally. Finally, the results of this evaluation are presented in Chapter 4.49Chapter 2Process Model Repair using DecompositionThis chapter describes the main contribution of this thesis: a modular technique for processmodel repair based on model’s decomposition. It is organized as follows.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
