As researchers, Jan Recker and I find it challenging to strike a balance between our efforts to meet the academic standards required by the wider research community and the demands regarding accessibility, relevance, timeliness and appropriateness instilled by the wider practitioner communities. We were happy to find that our blogging about research results inspires the BPM community to not only to take an interest in our research but also to critically assess this work and to post replies to it. We find this most welcome.
Our previous post on the frequency of BPMN construct usage has generated a passionate response by Bruce Silver who we know and respect as a very active contributor not only to BPM blogging in general but also to BPMN education and application specifically. Bruce makes many good points in his post and raises a number of interesting challenges. However, on some accounts we disagree with a number of the inferences he draws, so we want to clarify some aspects of our original post.
First of all, the paper and post are the result of a joint research effort between Jan Recker (QUT Brisbane) and myself (Stevens Institute of Technology), which we have stated. Jan and I started working together due to the complementary nature of our interests - standards in BPM (myself) and practical usage of modeling methods (Jan). Our study has been motivated by the fact that we know so very little about how standards such as BPMN are actually used - as opposed to what vendors, consultants and trainers think how they should be (or might be) used. - and this is what we try to explore and understand. The post and the related study are but one snapshot of our combined research. Agree with our results or not, but please give credit where credit is due. Jan is one of the most prolific researchers on BPMN; and it would be unfair to ignore his substantial contribution to this research.
We have outlined our research method in great detail in the full version of the BPMN paper, the PDF of which has been linked to from the original post. If you missed it, click here. Of course, we could have written a great deal more about the mode of analysis, but let’s be frank: how many blog readers would want to see this information in the post? (let us know if you do!) And for those of you taking an interest in research methods - both Jan and myself are more than happy to discuss the ways in which academic research is conducted. More than welcome.
We started with a simple question: BPMN is divided into a core and an extended set of constructs - does this separation hold in practice, or are there other common subsets (dialects or creoles) that can be found in practice? If there is such a common subset, we would expect a sizable number of models to share it. We found no evidence of a larger common core. Only 6 model pairs out of the 126 models used similar BPMN subsets (i.e. there were 6 subsets shared by 2 models).
We looked at the similarity among all subsets by coding the occurrence of symbols as a 50-bit string and computing the pairwise Hamming distance. On average 7 symbols differed between the BPMN subsets, and since the average model used only 9 symbols that makes the true common core very very small.
We performed a hierarchical cluster analysis on the models, trying to find the constructs that were used in groups. Indeed, several well-defined clusters emerged from this analysis: Basic Modeling Constructs, Annotations and Explanations (which include the blank XOR Gateway - not something we expected), Organization Modeling Constructs, and Control Flow Refinements. Users that move beyond these clusters seem to add individual constructs as needed, but in a rather random fashion.
Whether the 126 models we gathered are representative to all BPMN uses is a good question. Of course, we don’t claim this to be the case and we are in fact expanding our collection of models (hey Bruce, want to send us some of your seminar models?). However, so far our results have proven stable. We spend a great deal of our time with organizations using BPMN and we can assure you upfront - this is indeed indicative of how people use BPMN.
Bruce likens a frequency count of BPMN symbols to a character count in a document. We disagree - BPMN symbols are more like words, since they have semantics and are governed by formation rules. There are no formation rules at the character level in most languages. One could liken the frequency count to the frequency with which words in the English language are used - and that provides a much more useful metric than a character count. Linguists talk about the difference between an active and a passive vocabulary - words that we use versus words that we understand. It is possible that the use of BPMN is emerging along the same lines - a modeler might understand many of the symbols, but will frequently restrict him or herself to a more limited subset. To illustrate this: You may understand many entries in Merriam-Webster’s dictionary of the English language, but you do not use them frequently (or at all).
Do the models we collected have errors? Absolutely. Some of them we find useful in modeling courses - to show the types of errors usually made in practice. Our intention was not to analyze perfect BPMN models - we find those in every training course and in tool documentations, etc. The BPM reality looks different. Our intention was to analyze the current practice of BPMN modeling, not the indended application of the language. English speakers abuse their language - I know I do - but that does not mean that their sentences are meaningless.
Turning to some of the conclusions we draw from our research, we would like to clarify some aspects: What we call ‘the real core set of BPMN’ is what our analysis showed to be the most frequently used BPMN symbols found in the models considered. This does not mean we imply this set to be the core set of BPMN to be used by everyone. Rather, this is the minimal set of BPMN constructs actually used in practice so far. Is this set little more than flowcharting? Absolutely true. Absolutely.
But what does that tell us? People, and organizations, use BPMN for purposes similar to those organizations ten, twenty years ago that employed flowcharting - they want to describe their operations in simple, graphical terms. The process modeling efforts in most organizations at this stage are simply not advanced or mature enough to start specifying service-enable workflows with exception behavior in BPMN. No, most people use it simply for flowcharting.
What we conclude from this observation is that the ecosystem of vendors, consultants and trainers should be aware of this and should plan, manage and employ their efforts (be it tool development, BPMN training or modeling workshops) accordingly. We present a number of conjectures based on these observations, some of which appear to be troubling to Bruce. This is worrisome to us, we hope we can clarify this a bit more:
First, we see a great deal of training programs introducing the full BPMN specification to large number of stakeholders. Our results show, however, that most of this training is in fact only applicable to a small number of BPMN application areas. So we have to ask: Are there any tailored BPMN training programs? What should the ‘BPMN beginner’ course look like and how this body of knowledge then be extended by specialist courses? One of the suggestions we raise is indeed to start with the set of BPMN symbols that in fact are widely used in practice. Why? Because this would allow the BPMN beginner to instantly be able grasp, understand and use the majority of models in practice. Sure, (s)he would not yet be an expert, sure (s)he would not yet have learned about the benefits and expressive power of advanced BPMN. But (s)he can go out and leverage the knowledge instantly and make contributions. Without having to digest the complexity of a full-blown course. We do not imply that business users do not understand the more refined BPMN symbols, we have just found little evidence that they use them frequently.
Second, we suggest to tool vendors to rely more on empirical information about BPMN use when having to make trade-off decisions in BPMN support. Let’s face it - many BPMS do not support the full set of BPMN constructs. This makes sense, because if the system does not have the capability to execute the semantics of a specific construct (say, a transaction around a set of activities) then if would not make sense to allow a system analyst to draw this symbol. So which constructs can a vendor neglect initially and which need to be supported? We would argue that it is of best interest to vendors to focus on those constructs heavily used in practice. Why? Because this would give them access to the widest share of the market. Simple as that. This does not mean, that our suggestion is of a static nature. Of course not. Over time, full support should be given - and (relating to our previous conclusion) also BPMN users should learn the advanced features of BPMN. But organizations and tool vendors alike often face a need to achieve results very very fast. Which also means that releases are built and deployed that are far from finished.
Third, we think that our last conclusion was misread. Our intention is not to discredit the sizable development effort that went into the BPMN specification. More than 120 people participated in more than 120 interactions, be they face to face or conference calls. That’s a lot of BPM expertise leading to the current specification. We do not discourage advancement. We actually like BPMN’s advanced vocabulary. But have you asked end users what they think? Well, we did. Not only in this study but also in Jan’s large-scale BPMN usability studies we did find that users are in fact very troubled by the sheer number of, for example, event constructs. Are they used at a large scale? No. Do users understand their full capacity? Typically not. Why is this not at all reflected in BPMN development? That is exactly our point. Sure, our argument is a somewhat provocative statement. But if it helps to channel some attention to end usage, that’s fair by our standards.
We know a great deal about what BPMN can do in theory, how it is implemented in tools, how training programs (like Bruce’s) look like and even how we generate code from the diagrams and how the semantics can be tested and vigorously verified. But what do we know about how organizations engaged in BPM initiatives use it? Very little. Again, we were motivated by exactly this dearth of knowledge about real-life BPMN practice. Why? Because our own experiences with BPMN and with those organizations using it gave us this hunch that the theoretical usage (what vendors and consultants and trainers tell us) often has little to do with what the end users think or do (the practical usage). And why is it important to know what the end users think and do? Because it can help the researchers, vendors, consultants and trainers of this world to channel their attention and efforts to those problems real users face. Instead of the problems we think exist in practice.
We try to feed our empirical research back to the BPMN community - in the form of blogs, practitioner papers, or even directly by knocking on the door of OMG. Whether we are heard, and whether our findings have the type of impact we hoped is a different story. But we are always open for debate.
by Michael zur Muehlen (mzurmuehlen@stevens.edu) and Jan Recker (j.recker@qut.edu.au)
BPMN is the de facto standard for graphical process modeling. While there are other graphical languages to represent processes (EPCs, IDEF, Flowcharts, Petri Nets, among others), no other notation has seen such an uptake in such a short time as BPMN has. It is widely supported by both free and commercial process modeling tools, the WfMC has made XPDL 2.0 and 2.1 a de-facto persistency format for BPMN diagrams, and a large number of courses on modeling processes with BPMN are being offered.
Now, BPMN is a complex language. The current incarnation (BPMN 1.1) consists of 52 distinct graphical elements: 41 flow objects, 6 connecting objects, 2 grouping objects, and 3 artifacts. That’s a lot of vocabulary to learn, given that each graphical elements has meaning and rules associated with it. So what is the minimum subset of BPMN that a process modeler should know?The answer: Less than you think.
To answer this question we collected a large number of BPMN 1.0 diagrams (126 in total), from consultants, seminar participants, and online sources. We analyzed which BPMN symbols were actually used in these diagrams. The full version of our research, which we will present at the Conference on Advanced Information Systems Engineering in June, can be found here. But since this is an academic paper, here are the practical highlights of our study.
None of the diagrams we looked at used more than 15 different BPMN constructs, and none used less than 3. The models themselves contained considerably more elements, but a model with, e.g., 5 tasks connected by sequence flow was recorded as using the task symbol and the sequence flow symbol. The average subset of BPMN used in these models consisted of just 9 different symbols. That means that the average BPMN model uses less than 20% of the available vocabulary.
Figure 1 shows which construct we found across which percentage of the diagrams we collected.
Figure 1: Frequency distribution of BPMN construct usage
The results of our study are:
Only five elements (normal flow, task, end event, start event, and pool) were used in more than 50% of the models we analyzed. These, plus the data-based XOR gateway form what we call the common core of BPMN (marked in yellow in fig. 1).
Six additional elements were found in at least 25% of the models - gateways (parallel and unmarked XOR), lanes, text annotations, message flow, and start messages, we call these the extended core of BPMN (marked in green in fig. 1).
17 elements were used in less than 3 models - seven elements occurred in just two models, five in just one, and five elements were not used in any of the models we studied.
We then looked at the co-occurrence of BPMN symbols - i.e., are certain constructs used in combination, and how frequently? The combination of certain elements is mandated by the BPMN specification - you cannot use lanes without pools, or data objects without associations. But if there is a common subset used by many models, this would constitute a true “common core”. A detailed analysis revealed that BPMN elements fall into several well-defined groups. Figure 2 shows these groups as frames around the respective BPMN elements. The numbers within each frame represent the number of models (out of 126) that contain all elements within the frame.
Figure 2: Grouping of BPMN elements
Our findings are:
The common core of BPMN is very small. The subset of BPMN across the different models varied considerably. While nearly all models contain tasks and sequence flow, adding symbols to this set leads to a near exponential drop in models that share the (bigger) set of symbols. For example, while 65 models contain tasks, sequence flow, start and end events, only 25 also contain parallel gateways, and just 10 contain parallel gateways and data-based XOR gateways.
There are two types of BPMN modelers. While our sample is too small to explore this proposition in detail, we found anecdotal evidence that two groups of modelers use BPMN: Those who use pools and lanes to represent organizational responsibility for tasks, and those who use gateways to represent the control-flow rules of the process in detail. In other words, one group uses BPMN to specify inter-organizational settings (process choreography). Mostly, these users will be consultants or process analysts working on organizational (re-) engineering and process improvement. The other BPMN user group is leaning more towards workflow engineering (process orchestration). These users will likely be designers and analysts seeking to articulate precise flow conditions, for instance, in the context of workflow engineering or process simulation.
Implications
Our findings have implications for practitioners, software vendors, and standards makers alike.
Practitioners can begin studying the use of BPMN by focusing on the most commonly used symbols first, leaving more specialized and lesser-used constructs for those who need more specialized BPMN training (e.g. systems analysts).
Software vendors that are not supporting the entire BPMN vocabulary can assess what percentage of BPMN diagrams can be represented in their tool, and where enhancements should be made.
Finally, Standards-makers should review whether a more complete, but also more complex language is a desirable result of the standardization process. Creating BPMN took six years. How much time was spent on defining those seventeen symbols that we found are hardly used? And will the extensions of BPMN 1.1 entice users to expand their commonly used vocabulary, or will they go unused?
If you would like to learn more about this research, we encourage you to read the full version of our paper:
Michael zur Muehlen, Jan Recker. (Jun 16, 2008). “How Much Language is Enough? Theoretical and Practical Use of the Business Process Modeling Notation”, 20th International Conference on Advanced Information Systems Engineering (CAiSE 2008), Montpellier, France, June 16-20, 2008., Springer LNCS. Download (657 kb PDF)
The topic-du-jour in BPM circles is the handling of human and automated decision making in processes. Two major areas that intersect here are the management of business rules (such as “customers with more than 100,000 frequent flier miles receive priority treatment when flights are oversold“) and business processes (such as “rebook voluntary denied boarding customer“). There has been plenty of work done in both domains, but until very recently they did not talk to each other very much.
That has changed quite rapidly, as the business rules community realized that it needed some ways to represent the structured order of long-running decision-making activities (as typically found in workflows), and as the process community realized that modeling decisions and rules using activity networks, BPMN, or Petri Nets results in rather bloated and complex diagrams.
On the research side, my colleague Marta Indulska from the University of Queensland and I have studied the expressiveness of process and rule modeling languages using representational analysis (i.e. we used a formal ontology as a benchmark) and found that the combination of process modeling and rule modeling languages generally offers higher expressive power than either of these languages by itself. We found the combination of BPMN and SRML particularly useful, but since SRML is an abandoned effort we would recommend the combination of BPMN and SBVR. Our paper on this topic was presented at the VORTE’07 workshop and can be downloaded here.
In practice, you can do some process management with a rules engine. Natalie Glance and colleagues have written some intriguing papers on Generalized Process Structure Grammars in the mid-1990s that essentially allow the modeling of processes using a constraint language (saying things such as “the start of activity B must occur after the start of activity A”, which are difficult to express using languages such as BPMN).
In the same vein, you can handle quite a bit of decision making using graphical process modeling techniques, by building gateways into your processes. This way, your process diagram becomes (partially) a decision tree, with alternative pathways for different cases.
Coming from the process side of things, I found the formal logic and languages used in rules management standards such as RuleML rather intimidating. So when I was offered the opportunity to speak on the integration of rules and processes at the IIR BPM conference a few weeks ago, I tried to approach this topic from a pragmatic perspective:
Say you are a process modeler: How do you approach the topic of rules?
Most process models I’ve dealt with contain at least some aspects of decision making, typically found in splits (decision gateways). In particular two general types of decisions can be distinguished:
Decisions that affect the activities to be performed (Control Flow Decision). These types of decisions determine which process steps are appropriate for a given case (workflow instance). For instance, if you are dealing with a new customer and a large order you may want to perform a credit check, whereas you would skip this step if the customer is known to you. The decision in this case has an impact on the routing of the workflow instance.
We can distinguish some sub-cases in this scenario:
Single-criteria decisions where we only need to review one parameter
Multi-criteria decisions where we use decision tables or similar mechanisms to determine the case type
Similarly we can distinguish between:
Manual decisions, made by a human (e.g. when judgement is required)
Automated decisions that can be formalized
Decisions that affect the assignment of activities to performers (Assignment Decisions). These types of decisions determine who gets to perform a particular activity. For instance, a customer service representative may review an order up to $5,000, but above that value we want a manager to review the order. The review activity in both cases is identical, the difference lies in who gets to perform the work. The decision in this case has an impact on the assignment of the workflow instance.
So what is my point here? Whether you model these types of rules in your process modeling environment or not depends on your context, the availability of a separate rules management environment, and most critically, the frequency with which these rules change. There is no universally right or wrong way to manage the intersection of rules and processes. If your decision rules are as stable as your process, great, leave them in your BPM development environment. But if your business users want to manipulate the parameters, separate them from the process and handle them in a separate rules management environment. You’ll be glad you did.
My presentation from the IIR conference is available on slideshare.net (see embedded presentation below). And for some well-informed outside opinion you can refer to Sandy Kemsley’s timely blog post on the presentation.
I had the opportunity to spend a week in sunny Brisbane, Australia, to attend the 5th International Conference on Business Process Management, the largest academic conference in the BPM space. Hajo A. Reijers and I presented a paper on the use of biological mechanisms for task allocation, which are particularly useful in the emergency response domain (you can download the paper in the publications section here).Today I gave a two hour tutorial on the state of BPM standards, which was very well attended. The emerging landscape of integration, choreography, notation, and interchange standards in the BPM field are both confusing for practitioners and a fertile research field for researchers. Thankfully there are some efforts from standards groups to clarify the scope of their efforts and the impact of their specifications. I had a chance to talk to Karsten Ploesser of SAP, one of the co-authors of the BPEL4People draft specification, which is emerging as a very focused document, which will probably make it easier to understand and implement in practice - one key aspect that affects the adoption and diffusion of standards.Since I received a large number of requests for the presentation slides I made them available on slideshare. You can download the PDF of the presentation here and an audio recording is available here.
A little side comment about the do’s and don’ts of process discovery last month stirred up some interest, so Jan Recker and I got together and wrote a little piece.
Organizations have an unprecedented number of opportunities available to change and improve their processes. Advances in organizational research, mature supporting information systems, and complementary service offerings allow them to create compliant processes, improve their performance, simulate, automate, and monitor them. But one key challenge remains: You need to find your processes first.
During our BPM Day seminar on June 27th, part of the discussion focused on strategies for the identification of processes. Keith Swenson picked this up in a blog post here. We propose that process modelers seek out a rainy day scenario first, in order to discover as many process variants as possible early during their process modeling efforts. You can read our advice here:
zur Muehlen, M.; Recker, J.: Asking the Wrong Questions: Process Discovery on a Rainy Day. BPM.com, July 3rd, 2007. Available online at http://www.bpm.com/ [free registration required]
Entries (RSS)