I’ve been (slowly) reading Ten Steps to Complex Learning, by Jeroen J. G. van Merriënboer and Paul A. Kirschner.  The subtitle explains why: A Systematic Approach to Four-Component Instructional Design.

I read a lot about the death of instructional design, the end of training, and the New Jerusalem of learning that’s due any day.  Certainly a lot of superstition and nonsense gets daubed with the label “instructional design,” like a kind of cognitive Clearasil.  Still, I can’t help think that few people are going to learn to manage power-generation stations, conduct clinical trials, sell aircraft engines, or produce FEMA-acceptable flood elevation certificates solely through self-guided learning.

So I decided to plow through this book, which I’ve described with a bit of humor as being written in a language very much like English: the prose is dense, and very academic.  So far it’s worth the effort, and I’m going to summarize key parts here.

(Key part: something I pay enough attention to that I make a note on paper as I’m reading.  This is an ancient custom among my people.)

Van Merriënboer and Kirschner aren’t shy:

The fundamental problem facing the field of instructional design these days is the inability of education and training to achieve transfer of learning.

Which is something like AIG not being able to actually insure anything, isn’t it?

One point the authors make is that most complex skills require the learner to coordinate from a range of “qualitatively different constituent skills.”  That last phrase is important to them: not only is the whole of a complex task more than its parts, but the constituent skills are not parts of the larger task but aspects of it.  They’re not sub-skills, which you add together to make up the Big Skill.

Which, they argue, makes the analytic approach of many traditional instructional design approaches counter-productive.  For example, what they call the transfer paradox comes into play: the instructional methods that work best for isolated objectives often work poorly for integrated objectives.

To make that plainer: we spend too much time fiddling around with nice, clear, low-level objectives.  Then we lack time and money (and, perhaps, the will) to develop integrated learning.  Then we wonder why the training/learning function has such a dismal reputation.

But those isolated ones are what we tend to grab onto, because it’s easier to design around them, easier to create test items, and easier to cram them into an LMS (“Lessons Mean Simplicity”).

Learning to use the Amtrak reservation system is a complicated task, but maybe not all that complex.  Learning to act on traveler’s questions is also complicated.  Developing training for either set of skills is inherently less difficult than developing holistic training for an effective Amtrak reservation agent–but that’s what Amtrak’s really looking for.

The usual answer to the problem often seems to be “watchful waiting.” The performers go from training to the job, and we hope that their random encounters with reality end up filling the gaps.

Van Merriënboer and Kirschner want to grapple directly with such complex learning problems.  The model they advocate sees four main components to a learning blueprint:

• The learning tasks that someone needs to master.  (Strictly speaking, I’d say these are the on-the-job tasks which the person currently doesn’t know how to do, but it’s not my model.)
• The supportive information that comes into play when you’re working with skills that are performed differently from problem to problem.  These skills, which they call schema-based, benefit from things like mental models of the overall domain (e.g., pharma research) and cognitive strategies for working in that domain.
• The procedural information that guides those skills that are performed the same way from problem to problem.  This is the how-to knowledge (e.g., using the clinical trials database) that’s a routine part of the overall task.
• Part-task practice to strengthen and automate certain “recurrent constituent skills.”

Van Merriënboer and Kirschner argue that people can only perform certain constituent skills (which are aspects of the larger task, remember) if those people have a certain level of knowledge about the larger domain.  “Select an appropriate database,” as they point out, doesn’t make any sense if you don’t know what makes databases appropriate to the search you’d like to perform.

To foster integration and avoid compartmentalization, their model includes an emphasis on inductive learning: you as the learner work with specific problems so you build and improve mental models for the principles behind those specific problems.

…all learning tasks [should] differ from each other on all dimensions that also differ in the real world, such as the context…in which the task is performed, the way in which the task is presented, the saliency of the defining characteristics, and so forth.  This allows the learners to abstract more general information from the details of each single task.

That’s how we learn a great deal of what we know.  And, yes, a good deal of that happens informally, though I don’t see that as an argument for not trying to create learning situations when the informal can happen more predictably and more rapidly.

Related to this idea, the authors advocate always having learners work with whole tasks.  That might mean starting with simple cases or examples.  Other approaches include providing support (say, a process overview for the clinical-trial system) and guidance (a job aid for forming database queries).  They also make use of task classes, by which they mean categories of tasks.  In their ongoing database-query example, one task class has to do with performing searches when the concepts are clear, the keywords are in a specific database, when the search involves few terms, and where the result includes only a limited number of articles.

I’d call that the “clear, simple search” class.

You can imagine the other extreme: a poorly phrased request involving unclear concepts, with little knowledge of the appropriate databases, calling for complex search queries and producing large numbers of relevant articles which require further analysis.

How many task classes do you need?  That seems to depend on the range of variation between the Clear Simple Search class and the Nightmare Search class.

There’s a lot more going on; without intending to, I guess I’m starting another series.