The origins of cognitive load theory can be found in the results of a failed experiment. In the late 1970s, my research students and I were studying how people learned while solving problems (see Sweller, Mawer, & Howe, 1982). We were giving UNSW educational psychology undergraduates number transformation problems in which they were given a start number that had to be transformed into a goal number by finding the right sequence of moves. The only moves permitted were multiplying by 3 or subtracting 69. Problem solvers could use each of these moves as many times as they needed and in any order until they reached the goal number. For example, problem solvers had to convert 31 into 3 by successively multiplying by 3 or subtracting 69. (The answer is ×3, −69, ×3, −69.) Each move could be made just by pressing a key on a computer keyboard, so no mental arithmetic was involved. There were many problems, but each problem could be solved only by alternating the two possible moves a certain number of times.

The aim of the experiment was to see what factors would assist problem solvers in learning the rule, but we faced an immediate and inexplicable difficulty. As we sat there watching the participants in the experiment solving their series of problems, two things stood out. First, the problems were not very difficult, with most people solving them relatively quickly. Second, despite easily solving the problems, and despite the fact that successfully solving a problem meant that the alternation rule had been followed because alternating the two possible moves was the only way the goal could be reached, very few people were aware of the rule. They could solve up to sixteen problems, with most problem solvers remaining quite oblivious to the fact that they had alternated the two moves to solve every single problem. Looking at how people learned while solving problems was proving a quite futile exercise because all our evidence suggested that most of the participants in our experiment were learning very little. Either we had stumbled on a spectacularly dense set of experimental participants, a conclusion I was reluctant to accept because they had already demonstrated considerable sagacity by choosing to study educational psychology with me, or alternatively, everyone has difficulty learning while problem solving.

If learning and problem solving are incompatible, that incompatibility suggested we needed a new view of problem solving, because the field was embarking on a long excursion in which learning through problem solving was a basic assumption. The publication of Newell and Simon's (1972) book on problem solving inaugurated a flowering of the field. By studying people's cognitive processes while solving a problem and with the help of artificial intelligence programs, Newell and Simon were able to uncover some of the important mechanisms of problem solving. For about twenty years, problem solving became one of the central fields of cognitive psychology, with a large number of researchers engaged in studying the mental processes we engage in when solving a problem. Furthermore, since problem solving was central to education and training, descriptions by researchers of mental processes while students were solving a problem tended to be associated with advice from the researcher on how to design instruction. Most of that advice implicitly assumed that problem solving in one form or another was also the best form of learning.

While it is now quite clear that this advice was misguided, it is instructive to consider why such erroneous views could have been so influential. The proponents of the learning through problem solving recommendation collected a vast amount of data on the processes used by problem solvers. They avoided running controlled experiments in which learning was compared under problem solving as opposed to alternative conditions. No instructional recommendation should ever be accepted without that recommendation having been tested using controlled experiments in which the recommended new procedure is compared to currently used alternatives. Cognitive load theory has proved successful not only because of its reliance on a particular view of human cognition but also because no instructional recommendation generated by the theory has been offered without first being extensively tested using controlled experiments. All of the recommendations found in this book fall into this category. We should never accept instructional procedures that have not been tested in this fashion. Without controlled experimental testing, we face an unending list of instructional fads.

The failure of our problem solvers to learn anything useful about the structure of the problems we presented to them immediately led to the obvious question: How should we have taught them? In the case of the puzzle problems we were using, the answer was as obvious as the question. If you show people the alternating rule (×3, −69) rather than have them attempt to discover it, they will learn it immediately. As confirmed by both commonsense and controlled experiments, while the alternating rule may be hard to discover, it is trivially easy to learn.

Why is the rule hard to discover? While our language has altered somewhat over the years, from a cognitive load theory perspective, the explanation is straightforward. Solving a problem by searching for a solution places very heavy demands on working memory capacity. Those demands may result in a successful problem solution, but they do not leave sufficient working memory capacity to note which moves are appropriate for particular problem situations. Specifically, in the case of the numerical problems requiring alternation of the problem-solving moves, problem solvers devoted all of their working memory resources to working out which move was best at each point. They had no working memory capacity remaining to attend to the relations between moves. Once they made a move, they ignored both that move and all preceding moves and devoted all of their resources to working out what to do next. Without attending to the relation between moves, it is not possible to discover the alternation rule. The result was a successful solution with no realization that an alternating sequence of moves had been used.

While this result clearly had implications for the psychology of problem solving, in isolation, it could have no consequences for instruction. No findings using puzzles could have instructional implications. At this point, it was possible to either continue studying the psychology of problem solving as pure research or attempt to find applications for the finding by branching into instructional design. I decided that the finding had sufficient potential to risk attempting to replicate it using real educational materials rather than puzzle problems.

The first step was to find instructional areas where our puzzle problem findings might be relevant. Mathematics and science provided obvious examples. These areas had fairly stereotyped instructional procedures. A new area is introduced, one or two worked examples demonstrating how problems are solved using the new information are presented, followed by a relatively large number of problems for learners to solve. Since students are expected to spend a considerable time solving these problems, it is assumed that considerable learning will occur during this problem-solving process. If our puzzle problem results generalized to mathematics and science, this assumption may not be valid.

We tested two techniques that hypothetically should have been superior to conventional problem solving. The first required people to practice solving goal-free rather than conventional problems. For example, a goal-free geometry problem requires problem solvers to "calculate the value of as many angles as you can," rather than "calculate a value for Angle X." When calculating a particular angle, problem solvers must consider what they have now, consider the angle for which they must find a value, and search for a route between them. Working memory load is high. In contrast, when calculating the value of as many angles as possible, problem solvers merely look at their current problem state and calculate a value for any angle they can. Working memory load is substantially decreased. Our results strongly supported this hypothesis, with learners given goal-free problems performing much better on subsequent tests consisting of conventional problems, providing an example of the goal-free effect (Sweller, Mawer, & Ward, 1983).

The second technique we tested was the use of worked examples, as indicated in Chapter 8. We taught students how to solve simple algebra problems of the sort: (a + b)/c = d, solve for a. Then we gave a problem group lots of problems to practice while a worked example group were given the same problems with a worked solution for every second problem. On subsequent test problems, the worked example group did better than the problem-solving group, demonstrating the worked example effect (Cooper & Sweller, 1987; Sweller & Cooper, 1985). We suggested the superiority of the worked example group was due to cognitive load because they spent less time studying worked examples than the problem group spent solving the equivalent problems. Paas (1992) and Paas and van Merriënboer (1994) conclusively demonstrated that the effect was due to a reduced cognitive load for the worked example group using subjective ratings of task difficulty.

The obvious next step was to try the effectiveness of worked examples in other areas, such as geometry and physics. We ran exactly the same experiments using geometry and physics problems. To our surprise, the results indicated an utter failure of worked examples, with no evidence of superiority over problems. At that time, we had no idea why presenting learners with worked examples in algebra worked so well, while presenting geometry or physics worked examples was no better than solving the equivalent problems. It took considerable thought over several years to figure out why worked examples failed in some areas and, as often happens, the answer was staring us in the face.

As instructional professionals, our own working memory load is reduced by simply focusing on a particular instructional procedure such as the use of worked examples, rather than considering the cognitive processes underlying the instructional procedure. Assuming that practice problems should be transformed into worked examples is not cognitively taxing for us. Unfortunately, it is too easy and we need to analyze our instructional materials much more deeply. No simple technique is going to be universally applicable and, according to cognitive load theory, an instructional technique intended to reduce extraneous cognitive load will be effective only insofar as it does just that—reduce extraneous cognitive load. Algebra worked examples were effective, not just because they were worked examples but because they were worked examples that happened to be structured in a way that reduced extraneous cognitive load. Geometry and physics worked examples were ineffective because they were structured in a manner that failed to reduce extraneous cognitive load. We had structured our worked examples in a way that was conventional for their respective areas and, in the case of algebra, that structure coincidentally imposed a minimal, extraneous cognitive load. In the case of geometry and physics, the conventional structure of worked examples imposed a load on working memory that was no less than that imposed by solving problems. As a consequence, it made no difference whether we got people to study our worked examples or solve our problems.

How should we structure geometry or physics worked examples? The answer is—in a manner that reduces split attention. Chapters 4 and 8 provide examples. For example, when faced with a diagram and text that are unintelligible in isolation, learners should not have to search for referents. Searching for which parts of a diagram go with which parts of text requires working memory resources that consequently become unavailable for learning. We found that constructing worked examples that reduced or eliminated that search by physically integrating text into diagrams reinstated the worked example effect in geometry and physics and, in the process, demonstrated the split-attention effect (Sweller, Chandler, Tierney, & Cooper, 1990; Tarmizi & Sweller, 1988; Ward & Sweller, 1990). Eliminating split attention is, of course, important in all instruction, not just worked examples.

We had failed to generalize the worked example effect in a variety of areas until we considered cognitive load theory more deeply. That failure allowed us to use cognitive load theory to generate the split-attention effect. In many ways, that sequence of events provides a template for the history of cognitive load theory over the next fifteen years. We repeatedly found a new instructional effect or procedure, found the limits of that effect when attempting to generalize it to new conditions, and used cognitive load theory to generate a new effect. Over many years, we realized that a successful experiment almost never leads to either theoretical or practical advances. In our case, advances seemed to come from failed experiments. Successful experiments did little more than confirm what we already knew.

The redundancy effect grew from the split-attention effect in exactly this manner. We had found that learning was facilitated by integrating text with diagrams. We automatically assumed that integrating disparate sources of information would always be successful but, of course, instructional design is never that easy. The cognitive load implications of integrating text and diagrams matters, not the act of integration. It took us a few months to realize that, if text merely re-described a diagram, little was gained by integrating the sources of information. Integration was very successful when both sources of information were unintelligible in isolation and so had to be integrated either mentally or physically to reduce cognitive load. When the same information was merely repeated in a different form, cognitive load was increased by having to unnecessarily coordinate both sets of information, irrespective of whether they were integrated or not. The best way to reduce extraneous cognitive load was to eliminate the redundant version. Chandler and Sweller (1991) demonstrated the redundancy effect by showing that two forms of the same information resulted in less learning than one form.

It turned out that we were not the first to demonstrate the redundancy effect. That effect had been demonstrated, forgotten, and then demonstrated again on several occasions over many decades. Why was it forgotten on each occasion? For many people, the redundancy effect is counter-intuitive. Most of us intuitively feel that presenting learners with the same information in several different ways cannot be harmful and could be beneficial. In fact it is harmful, and cognitive load theory explains why it is harmful. If we have to unnecessarily coordinate multiple sources of the same information, scarce working memory resources are being used for activities unrelated to schema acquisition and automation, depressing learning. Demonstrations of a counter-intuitive effect without a proper theoretical explanation tended to be ignored and forgotten, and that may be why early examples of the redundancy effect had no impact on the field. It is to be hoped that the current, cognitive load theory explanation can establish the redundancy effect as a major weapon in the armory of instructional professionals.

The split-attention effect led indirectly to the redundancy effect as described above. It led directly to the modality effect. It has been known for some time that both auditory and visual working memory can be used simultaneously and that, in combination, the use of both processors increased the capacity of working memory to some extent (Penney, 1989). When faced with a diagram, which must be visual, and words, which can be either visual or auditory, it seemed to follow that audio/visual presentation with the words presented in auditory (that is, spoken) form should increase available working memory by transferring some of the visual working memory load to auditory working memory. Mousavi, Low, and Sweller (1995) tested the hypothesis that, under split-attention conditions (that is, where the two sources of information were unintelligible in isolation), presenting text in spoken rather than written form would be beneficial. Controlled experiments comparing geometry diagrams and written or spoken text demonstrated the superiority of the spoken text, providing an example of the modality effect.

The split-attention, redundancy, and modality effects proved strong, robust effects that could be easily demonstrated and so provided clear principles for instruction. The only problem was that, mixed with some large experimental effects, we kept obtaining evidence that under some conditions there was no effect at all. Again, these failures forced us to return to basic cognitive load theory to provide reasons that could be used to generate new instructional principles. In the process, cognitive load theory underwent a massive expansion.

To this point, we had only been concerned with extraneous cognitive load, which is cognitive load caused by instructional procedures. Many of the cognitive load principles discussed in Part 2 of this book are examples of extraneous cognitive load because they are under the control of the instructional designer. There are other forms of cognitive load—with intrinsic cognitive load being an important example.

Here's how the concept of intrinsic cognitive load developed. In the early to mid-1990s, we realized that some effects, such as the split-attention, redundancy, and modality effects, could not be obtained with some materials. We needed an explanation. We eventually discovered that the effects invariably failed when the nature of the material was such that it could be processed in working memory one or two elements at a time (Sweller, 1994; Sweller & Chandler, 1994). Learning to translate some of the nouns of a foreign language provides an example. You can learn each noun without reference to any of the other nouns, and so working memory load is naturally low. In contrast, the natural structure of some material is such that working memory load is high during learning. In order to understand such material, you needed to process many elements simultaneously because they interacted (for example, learning to deal with an equation or formula). Using this high element interactivity material, split-attention, redundancy, and modality effects could be readily obtained. In other words, complex material with many interacting elements gave the effects, but simple material with very few or no interacting elements did not.

The reason instructional effects could be obtained using high but not low element interactivity material was that if the material included many interacting elements, it imposed a high intrinsic cognitive load—intrinsic because it was not determined by what the instructor did, but by the nature of the material. If a high extraneous cognitive load due to instructor activity was added to a high intrinsic cognitive load due to high element interactivity, we got the various effects. If intrinsic cognitive load was low due to low element interactivity, it hardly mattered what the instructor did because working memory was not overloaded. In other words, with a low intrinsic cognitive load, extraneous cognitive load didn't matter. We called this effect the element interactivity effect.

At that time, we assumed intrinsic cognitive load was immutable. It could not be varied because it was "intrinsic" to the material. Only extraneous cognitive load due to instructional design could be varied. With Pollock, Chandler, and Sweller (2002), we realized that there had to be ways of reducing intrinsic cognitive load; otherwise very complex material could never be learned. We had to modify the theory to say that you can reduce intrinsic cognitive load but you cannot simultaneously maintain full understanding. You can eliminate and reduce some of the interacting elements to allow working memory to handle the material. For example, in a computer application, you may omit explanations and just tell learners what steps to follow. Those steps may be easily processed in working memory. At that point, understanding won't occur, but once the reduced material has been learned it can be put together with the omitted information to give understanding. In that sense, intrinsic cognitive load is to some extent under the control of the instructor. Learning can be facilitated by reducing the number of interacting elements and only reintroducing them later when the essential elements have been learned.

For these reasons, the split-attention, redundancy, modality, and other effects are defined as being due to extraneous cognitive load because they are under the full control of the instructor. Intrinsic cognitive load is not. It is intrinsic to the material being taught. The instructor will need to devise instruction to take account of intrinsic cognitive load by omitting some of the interacting elements, but cannot alter it except in such artificial ways and with an initial loss of understanding. Nevertheless, by presenting material with some interacting elements initially omitted, learning can be facilitated. We called the effect the "isolated/interacting elements effect."

A third form of cognitive load, germane cognitive load, was discovered after cognitive load theory attracted international interest. That interest first arose in Europe. Until the early 1990s, cognitive load theory was exclusively studied at the University of New South Wales in Sydney. While we were a relatively large group, we were the only group. That situation began to alter when Jeroen van Merriënboer and his then student, Fred Paas, working in Holland, began using the theory.

Paas and van Merriënboer (1994) found that if they gave learners worked examples that differed considerably in variability, cognitive load was increased compared to worked examples that were all very similar. Nevertheless, despite the increase in cognitive load, high variability worked examples resulted in better learning than low variability examples, giving the variability effect. Clearly, this was a different type of cognitive load from the more commonly studied extraneous and intrinsic load. They labeled this form of cognitive load "germane" cognitive load because it was a load that was germane to schema acquisition and automation. In effect, the aim of reducing extraneous cognitive load is to free working memory capacity for germane load. If the only consequence of reducing extraneous cognitive load is to reduce the mental work that learners do, it will not result in improved learning.

The introduction of germane cognitive load was not the only advance provided by Paas and van Merriënboer (1994). Along with Paas (1992), they provided a very timely technique for directly measuring cognitive load. In the early 1990s, cognitive load theory was beginning to be noticed for the first time, but it was also beginning to be criticized. Much of this criticism was ideological in nature. Instructional theory was going through one of its frequent fads based on ideology, rather than data or on any coherent conception of human cognitive architecture. This particular fad insisted that instruction should never directly provide students with information; rather, they should discover it themselves. The name of the movement kept altering as it failed to find credible data, but you may recognize discovery learning, problem-based learning, enquiry-based learning, or constructivism. By the early 1990s, this movement was at its height, and the last thing its adherents wanted was someone from the wrong side of the world telling them that learning was facilitated by directly instructing learners and ensuring that the instruction reduced unnecessary mental activities.

Up to that point, there was a large hole in the empirical evidence for cognitive load theory, and that hole could be exploited by anyone wishing to attack the theory. While we could and had used the theory to generate lots of novel instructional techniques, our evidence that the effects were caused by cognitive load rather than some other factors were indirect at best. A reduced learning time was the most frequently used evidence. Paas (1992) and Paas and van Merriënboer (1994) demonstrated the worked example effect, but in addition used subjective measures of task difficulty to directly measure cognitive load. They not only found that the use of worked examples facilitated learning but also found that learners experienced a reduced working memory load as a consequence of studying worked examples. We immediately began using their techniques at UNSW to demonstrate that cognitive load was indeed the reason for the cognitive load effects.

One of the other findings provided by our Dutch colleagues at this time was the example completion effect. Rather than giving people full worked examples to study, they gave them partially completed problems that had to be completed by the learner. Paas (1992) and Paas and van Merriënboer (1994) found this technique equally as effective as worked examples and better than full problems, thus providing an example of the example completion effect. This effect was subsequently used by Renkl and Atkinson (2003) to demonstrate the guidance fading effect. They provided learners with full worked examples initially. As learner expertise increased, those full examples were replaced by partially completed examples. With additional expertise, the partially completed examples were replaced by full problems. Some new findings at UNSW indicated the need for this guidance fading procedure.

All of the effects described so far were demonstrated using novices. In the late 1990s we discovered that, as expertise increases, the effects gradually disappear. With further increases in expertise, they reverse (see Kalyuga, Ayres, Chandler, & Sweller, 2003). Again, cognitive load theory can be used to explain why. Consider worked examples. For novices, they are needed to demonstrate how a problem or class of problems is solved. As expertise increases, learners may still need to practice, but worked examples may now be redundant. Unnecessarily studying a worked example may impose a greater extraneous cognitive load than simply solving the problem, resulting in a reverse worked example effect. At this point, problem-solving practice may be more useful than studying worked examples (Kalyuga, Chandler, Tuovinen, & Sweller, 2001). In other words, worked examples need to be faded through completion problems to full problems for maximum learning efficiency.

At this time, many other researchers began to use cognitive load theory. In the United States, Richard Mayer incorporated the theory into his cognitive theory of multi-media learning (Mayer, 2001). He, along with Roxana Moreno (Mayer & Moreno, 2003) focused the use of cognitive load theory principles on instruction that used words, pictures, and sound. Several researchers in Germany began to use the theory in their experiments. The work of Alexander Renkl, along with Robert Atkinson (the latter in the United States) in combining the worked example, completion, and expertise reversal effects to develop the guidance fading effect, has been described above. Roland Brunken, Detlev Leutner, and Jan Plass (the latter in the United States) conducted some very substantial work on the use of secondary tasks as an alternative to subjective rating measures of cognitive load (Brunken, Plass, & Leutner, 2003). Peter Gerjets along with Katharina Scheiter and Richard Catrambone (the latter in the United States) investigated instructional techniques for handling a large intrinsic cognitive load (Gerjets, Scheiter, & Catrambone, 2004). All of these researchers have done more than merely use cognitive load theory. As a consequence of their work, the theory has altered and developed.

Currently, many lines of work are being pursued within a cognitive load theory framework. Summaries of some of that work can be found in recent special issues of journals devoted to the theory (Educational Technology Research and Development, 2005, 53[3]; Educational Psychologist, 2003, 38[1]; Instructional Science, 2004, 32[1–2]; Learning and Instruction, 2002, 12[1]). At UNSW, we are pursuing many different lines of work. Three lines that I am closely involved with are (1) finding measurement devices that can be used to determine what sort of information should be provided to learners; (2) looking at the effect of asking learners to imagine concepts or procedures; and (3) using our instructional findings to further develop our knowledge of human cognition, especially the reasons why human cognitive architecture evolved in its particular way.

From its beginnings as an attempt to understand why problem solvers learn little while solving problems to its current concerns with how our cognitive architecture is structured, cognitive load theory has continually based its instructional recommendations on the outcomes of controlled experiments. That emphasis on research into human cognition to provide theory and controlled experiments to provide data has led to the development of the instructional principles described in this book. While most of those principles are concerned with learning, a limited working memory affects all of our activities when dealing with novel information, including perceiving information and understanding instructions. The limitations and strengths of our cognitive architecture affect all of our cognitive activities. Ultimately, beyond the science of cognition and instruction, the usefulness of a cognitive load theory based approach will be determined by practitioners in the field.