Based on the Decision Model and Notation (DMN), decision logic can be outsourced from process models to standardized decision models. This requires considerations of the sound integration of the two types of models. In this paper, we describe an open-source tool that checks the soundness of such decision-aware process models, inspired by the traditional workflow soundness notion. This includes ensuring that the process can continue after a decision has been taken, and that all activities following the decision can be executed. Our tool is scalable and its relevance has been established based on two empirical evaluations.
from now on. Based on some user info the decision task determines a set of ads that are relevant for the customer. Afterwards, depending on the outcome of the decision, measures such as displaying warnings about alcohol must be taken, or if only toys should be advertised the user is redirected to a designated webpage for children. We call these fragments of process models decision fragments, defined by their structure of beginning with a decision task, followed by a split gateway, followed by two or more tasks. The tasks are enabled via the conditions of the split gateway’s outgoing edges, which refer to the outcome of the decision. The DMN decision table associated with the decision task Determine ads is shown in Fig. 2. This table is the top-level decision table of a DMN decision model, consisting of just this table with two inputs, Age and Has Children. The corresponding DMN decision requirements diagram is not shown for simplicity.
Complementing process models with decision models requires considerations about
the correctness of the resulting decision-aware process models. The presented tool checks
two behavioral criteria that must be fulfilled for a correct integration, namely decision
deadlock freedom and dead branch absence [
The tool to check the soundness of decision-aware business processes is implemented in
dmn-js, a DMN decision table editor developed by Camunda1. Parts of our
implementation rely on functionality provided in a tool to verify DMN decision tables [
In the following, we will apply the tool to the example from Section 1. Fig. 3 shows the interface of the tool after the process model was imported to the decision table editor via the import BPMN button and the check soundness button was clicked to check for the two behavioral soundness criteria. The results are presented in the table Analysis results. First of all, the output set of the decision table is computed and displayed. This is defined to be the set of outputs the table can produce given the domains of the input variables, the rules, and the hit policy. The table in the example has four rules that operate on an integer and a boolean input. Furthermore, it is a multi-hit table with a rule order policy, denoted by the letter R in the upper left corner. Consequently, if two rules rules match for the same input, the output of both rules will be returned as a list ordered by the order of the rules. For example, only rule 4 matches for the input (10, f alse) (or (10, true)), so the only output will be T oys. However, for input (13, f alse) both rules 2 and 4 match, such that the outputs of both rules are returned in a list: [V ideogames, T oys]. 3
After the entire table output set was computed, the behavioral criteria are checked. In the example in Fig. 3, the decision deadlock freedom criterion is violated because no edge condition of the decision fragment evaluates to true for the output [V ideogames, T oys]. All other outputs fulfil at least one edge condition. For example, the output T oys perfectly matches the lower branch condition and the upper condition is satisfied for the last two outputs because they contain the element Alcohol.
To get rid of the error, we could now either change the table so that every output is covered by an edge condition or adapt the decision fragment so that the uncovered output is accounted for. For instance, we could try to add another branch with the condition Ads == V ideogames and check soundness again. This will lead to the result displayed in Fig. 4. Unfortunately, the output [V ideogames, T oys] is still uncovered. Additionally, by adding the new branch we violated the dead branch absence criterion since the condition Ads == V ideogames is actually unreachable. There is no output of the decision table that is exactly equal to V ideogames. To eliminate both errors the branch condition should be changed to Ads contains V ideogames. In this way, the uncovered output will match this condition, which in turn is not unreachable anymore. In this section, we discuss three aspects regarding our tool. First, we report our findings from two empirical studies we conducted regarding the practical relevanve of our notion of soundness of decision-aware business processes. Second, we discuss the scalability of our algorithm. Lastly, we describe our tool’s maturity.
In one study regarding the practical relevance of our notion we asked participants of an online course on process and decision modeling to design a BPMN process and an associated DMN decision model given a textual description of a billing process. We received 70 valid submissions, 43 (61%) of which violated at least one of the behavioral criteria. In another study, we analyzed a process management project of a large German insurance company that serves to handle invoices of insured customers automatically. This project contains 86 decision-aware business processes. In 26 (30%) of those we detected violations of our soundness criteria.
With respect to scalability, our implementation is divided into two parts, one for
computing the output set of a decision table, and one for checking the two behavioral
criteria. The former part extends the code for finding overlapping rules in a tool for
DMN decision table verification [
Our tool is able to check the sound integration of decision fragments with toplevel decision tables with any number of output variables (a DMN decision table may have more than one output variable column) and any kind of hit policy. Regarding the decision fragment, its split gateway can be an exclusive, inclusive or complex gateway. Furthermore, the edge conditions can be made up of conjunctions or disjunctions of unary tests. Each unary test is an expression that compares the output of the table to another value. For the comparison, in case of single value outputs the usual arithmetic comparison operators appropriate for the type of the variable can be used. In case of list outputs, equality and containment operators are available. In the end, the result of the comparison will either be true or false so that it can be easily decided whether or not an output is covered or a branch is reachable. Note that we assume that a decision table has finitely many rows and therefore also finitely many outputs. Similarly, a decision fragment is assumed to have a finite number of branches each consisting of finitely many unary tests. Therefore, our soundness check is guaranteed to terminate.