No two learners are identical. No individual learner is a perfect match for the one imagined by the learning designer. Thus, any learning content design—no matter how perfect—is to some extent wrong.
How wrong remains unknown until well after a learner completes it. Practically, though, that depends in large part upon how wrong it is at the time a learner engages with it. After completion, the negative repercussions could be:
- The learner still has unresolved knowledge/skill gaps,
- The time required to finish was longer than necessary, and
- The experience was less engaging than it could have been.
(Yes, that's the "effective/efficient/engaging" trinity.)
These negatives were bad enough before 2020. But now, in the shadow of COVID-19? With the pace of change we expect to see from 2022 onwards? We should prepare now for a future of constant upskilling and reskilling. Our learning & development content has to be less wrong than it is now.
So, how can we reduce this wrongness?
First, we must make sure our content targets legitimate performance goals. After that, another practical approach seems to be adaptive learning, i.e., learning content that adapts to the learner as it’s delivered. (Adaptive performance support could be another good approach, but won’t be discussed in this article.)
Getting real
In the real world, if a learner interacts with a human tutor, adaption happens naturally. The tutor asks questions and/or observes the learner, evaluates the learner’s activities and responses, and adjusts as needed.
Adaptive learning courseware must do the same. It may or may not adapt as well as a good human tutor, but compared to standard eLearning it’s infinitely variable (and complex).
There are many ways to look at that variability and reduce the apparent complexity. Here’s a useful 4 Characteristic Model of Adaptivity that can help.
- Adaption Priorities: What are the expected benefits from adapting?
- Adaption Modes: What forms or patterns will the adaptivity take?
- Adaption Bases: What information is used to make the adaption decisions?
- Adaption Targets: What aspects of the content will be adaptive?
The table below shows the model, along with some basic choices for each characteristic.
4 characteristic model of adaptivity
|
Adaption Priorities (Reasons/Benefits) |
Adaption Modes (Structures/Patterns) |
|
Reduce training time Minimize knowledge/skills gaps Increase acceptance/motivation Develop/promote advanced proficiency Maximize value of training time/effort |
Course sub/supersets for different audiences Content bypass Content remediation Starting point selection Learning path macro/micro adjustment |
|
Adaption Bases (Topic / Learner Characteristics) |
Adaption Targets (Instructional Elements) |
|
Topic intrinsic difficulty/complexity Topic domain/level Learner knowledge/experience Learner abilities/preferences Learner performance/achievement level |
Depth/breadth/pace/sequence Presentation mode (text, graphics, video, etc.) Problem/exercise type and difficulty Learning supports and feedback Degree of learner autonomy/choice |
You can select one or more items from each box to define a type of adaption. Here’s an example (actually, it’s two examples).
- Adaption Priority: Reduce training time
- Adaption Mode: Content bypass
- Adaption Basis: Learner performance/achievement level OR knowledge/experience
- Adaption Target: Depth/breadth/pace/sequence
If applied to eLearning, what do these look like in actual practice?
- The learner is offered a pre-test. If they pass, they get to bypass most of the content. If not, they just take the course as usual. In our prototypes, learners that took the pre-tests reduced their in-course times by about 40%.
- The learner is notified that their training history allows them to bypass some material. If they accept, they get to skip several different chunks of the course. In our prototype, learners that accepted were able to reduce their in-course times by 30–40%. (Depending on the details of their training histories.)
These are pretty basic levels of adaption, and they still led to significant benefits. But this begs the question: What are some other levels of adaptivity in eLearning?
Also, how would those two examples rank relative to other forms that are possible? Inspect the following table and try to place them. (Fractional levels are allowed.)
The levels of eLearning adaptivity
|
Level & Name |
Minimum Input |
Minimum Output/Action |
Notes |
|
0. Non-adaptive |
N/A |
N/A |
Feedback for individual questions or exercises might vary but the content of the next Screen, Slide, or Chunk (SSC) is always the same for all learners (excluding the effects of personalization that alters style or appearance). |
| 1. Simple bypass |
One of: pre-test score, job role, learner history, etc. |
Option to bypass larger of: 4 full SSCs OR 50% of a full module |
Stated minimums are intended to screen out trivial or inconsequential bypasses. |
|
2. Branching/ conditional control |
Single output from a just-completed question, exercise, etc. |
Next full SSC selected from at least two substantially different choices |
Intro and sum-mary SSCs need not be selected dynamically. Otherwise, the selected item can be less than a full SSC as long as it’s instructionally relevant. The most common expres-sions of this level are branching scenarios or branching programmed instruction. |
|
3. Model-based control |
More than just a single output from a just-completed question, exercise, etc. |
Next full SSC selected from at least two substantially different choices |
Intro and summary SSCs need not be selected dynamically. Otherwise, the selected item can be less than a full SSC as long as it’s instructionally relevant. The selection model is more complex than a simple Boolean logic check of one or more data points. It could be an inference engine, a feedback control model, an implementation of item response theory, a machine-learning model, etc. |
|
4. Dynamic/ responsive content generation |
More than just a single output from a just-completed question, exercise, etc. |
Next SSC is generated dynamically in real-time |
To the learner, this level would “feel” like interacting with a competent human tutor (a tutor that’s sane and doesn’t go off on weird, potentially inappropriate tangents). As of late 2021, an instructionally reliable implementation of this level seems almost possible. You’d start with a model-based controller similar to Level 3. Instead of selecting from pre-existing SSCs, though, use something like a transformer-based deep-learning engine (the best-known of these currently is probably OpenAI’s GPT-3). |
|
X. Magic.*
* Clarke’s 3rd Law: Any sufficiently advanced technology is indistinguishable from magic. |
One learner |
One performer |
Level X is essentially fantasy or science fiction. Is it a magical spell or incantation? Is it a mind meld or direct neural download? Does it involve a charmed amulet or a brain implant? Who knows?! In the future, if some new technology (or actual magic) enables adaptivity well beyond the highest-numbered level, add it to the table and assign it the next number in sequence. |
If you said the two examples were Levels 1 and something between 1 and 2, respectively, good job.
Many people are already doing some form of Level 2 adaption, and Level 3 (or slightly higher) is becoming more common all the time. (We have our own Level 3 model deployed and are using it in small ways now.)
Prototypes for Level 4 probably exist right now; I suspect there may be a viable commercial offering by 2025.
As for Level X… if it exists at all, it’s deep inside some secret underground lab somewhere. Don’t ask too many questions or you might end up as a unitard-wearing test subject with mysterious head bandages and no memory of how you got wherever you are.
Getting practical
When should you opt to design an adaptive course? Some cases are pretty obvious, but others aren’t. Choose your projects wisely because design/development costs can rise quickly as complexity increases.
Here are some criteria to help you decide.
Practical criteria for considering adaptive learning
- The course is required periodically (annually, biennially, etc.)
- The learner population is highly varied (knowledge, experience, etc.)
- The content spans a broad and/or deep cognitive/technical range
- The learning objectives cross over with other elements of the curriculum
- Using adaption could yield significant cost savings
- Basic adaption is an easy retrofit
The dominant eLearning authoring tools are capable of producing adaptive learning, but they don’t necessarily make it easy. Out of the box, you’ll be limited to Level 2 if you don’t want to do a lot of custom JavaScript coding.
The outlook is much better if you’re using one of the many cloud-based systems with built-in adaptive features. If you are, then you should definitely go adaptive often. Even then, however, you’ll still want to evaluate what level of adaptivity to use (if you have options) and how adaptive the course should actually be. (The increased design/development cost is still a factor, although likely mitigated somewhat by the features of the system.)
Closing
The future of business looks challenging but adaptive learning can help our learners prepare. However, adaptive learning is unlikely to become the mainstream standard unless more people start demanding it. Tell your authoring tool vendor that you want adaption-enabling features now if you don’t have them already.
References
W. J. WILSON, B. GALEMA, G. HAVERLUCK, “Adaptive Computer Based Training – Explorations, Experience, and Future Directions”, Proc. Conf. on Nuclear Training and Education: A Biennial International Forum (CONTE 2021), online (virtual), February 9–11, 2021, p. 15, American Nuclear Society (2021).