Permission to reprint or excerpt is granted only if the
following line appears at the top of the article:
ANTIC PUBLISHING INC., COPYRIGHT 1986. REPRINTED BY PERMISSION.
Column #8—User Interfaces, Homily #1
Professional GEM, by Tim Oren
In response to a number of requests, this instalment of ST PRO GEM will be devoted to examining a few of the principles of computer/human interface design, or "religion" as some would have it. I'm going to start with basic ergonomic laws, and try to draw some conclusions which are fairly specific to designing for the ST. If this article meets with general approval, further "homilies" may appear at irregular intervals as part of the ST PRO GEM series.
For those who did NOT ask for this topic, it seems fair to explain why your diet of hard-core technical information has been interrupted by a sermon! As a motivator, we might consider why some programs are said by reviewers to have a "hot" feel (and hence sell well!) while others are "confusing" or "boring".
Alan Kay has said that "user interface is theatre". I think we may be able to take it further, and suggest that a successful program works a bit of magic, persuading the user to suspend his disbelief and enter an imaginary world behind the screen, whether it is the mathematical world of a spreadsheet, or the land of Pac-Man pursued by ghosts.
A reader of a novel or science fiction story also suspends disbelief to participate in the work. Bad grammar and clumsy plotting by the author are jarring, and break down the illusion. Similarly, a programmer who fails to pay attention to making his interface fast and consistent will annoy the user, and distract him from whatever care has been lavished on the functional core of the program.
Before launching into the discussion of user interface, I should mention that the general treatment and many of the specific research results are drawn from Card, Newell, and Moran's landmark book on the topic, which is cited at the end of the article. Any errors in interpretation and application to GEM and the ST are entirely my own, however.
We'll start right at the user's fingers with the basic equation governing positioning of the mouse, Fitt's Law, which is given as
T = I * LOG2( D / S + .5)
where T is the amount of time to move to a target, D is the distance of the target from the current position, and S is the size of the target, stated in equivalent units. LOG2 is the base 2 (binary) logarithm function, and I is a proportionality constant, about 100 milliseconds per bit, which corresponds to the human's "clock rate" for making incremental movements.
We can squeeze an amazing amount of information out of this formula when attempting to speed up an interface. Since motion time goes up with distance, we should arrange the screen with the usual working area near the center, so the mouse will have to move a smaller distance on average from a selected object to a menu or panel. Likewise, any items which are usually used together should be placed together.
The most common operations will have the greater impact on speed, so they should be closest to the working area and perhaps larger than other icons or menu entries. If you want to have all other operations take about the same time, then the targets farthest from the working area should be larger, and those closer may be proportionately smaller.
Consider also the implications for dialogs. Small check boxes are out. Large buttons which are easy to hit are in. There should be ample space between selectable items to allow for positioning error. Dangerous options should be widely separated from common selections.
Anyone who has used the ST Desktop for any period of time has probably noticed that his fingers now know where to find the File menu. This phenomenon is sometimes called "muscle memory", and its rate of onset is given by the Power Law of Practice:
T(n) = T(1) * n ** (-a)
where T(n) is the time on the nth trial, T(1) is the time on the first trial, and a is approximately 0.4. (I have appropriated ** from Fortran as an exponentiation operator, since C lacks one.)
This first thing to note about the Power Law is that it only works if a target stays in the same place! This should be a potent argument against rearranging icons, menus, or dialogs without some explicit request by the user. The time to hit a target which moves around arbitrarily will always be T(1)!
In many cases, the Power Law will also work for sequences of operations to even greater effect. If you are a touch typist, you can observe this effect by comparing how fast you can enter "the" in comparison to three random letters. We'll come back shortly to consider what we can do to encourage this phenomenon.
Just as fingers are the way the user sends data to the computer, so the eyes are his channel from the machine. The rate at which information may be passed to the user is determined by the "cycle time" of his visual processor. Experimental results show that this time ranges between 50 and 200 milliseconds.
Events separated by 50 milliseconds or less are always perceived as a single event. Those separated by more than 200 milliseconds are always seen as separate. We can use these facts in optimizing user of the computer's power when driving the interface.
Suppose your application's interface contains an icon which should be inverted when the mouse passes over it. We now know that flipping it within one twentieth of a second is necessary and sufficient. Therefore, if a "first cut" at the program achieves this performance, there is no need for further optimization, unless you want to interleave other operations. If it falls short, it will be necessary to do some assembly coding to achieve a smooth feel.
On the other hand, two actions which you want to appear distinct or convey two different pieces of information must be separated by an absolute minimum of a fifth of a second, even assuming that they occur in an identical location on which the user's attention is already focused.
We are able to influence the visual processing rate within the 50 to 200 millisecond range by changing the intensity of the stimulus presented. This can be done with color, by flashing a target, or by more subtle enhancements such as bold face type. For instance, most people using GEM soon become accustomed to the "paper white" background of most windows and dialogs. A dialog which uses a reverse color scheme, white letters on black, is visually shocking in its starkness, and will immediately draw the user's eyes.
It should be quickly added that stimulus enhancement will only work when it unambiguously draws attention to the target. Three or four blinking objects scattered around the screen are confusing, and worse than no enhancement at all!
Both the information gathered by the eyes and movement commands on their way to the hand pass through short-term memory (also called working memory). The amount of information which can be held in short-term memory at any one time is limited. You can demonstrate this limit on yourself by attempting to type a sheet of random numbers by looking back and forth from the numbers to the screen. If you are like most people, you will be able to remember between five and nine numbers at a time. So universal is this finding that it is sometimes called "the magic number seven, plus or minus two".
This short-term capacity sets a limit on the number of choices which the user can be expected to grasp at once. It suggests that the number of independent choices in a menu, for instance, should be around seven, and never exceed nine. If this limit is violated, then the user will have to take several glances, with pauses to think, in order to make a choice.
The effective capacity of short-term memory can be increased when several related items are mentally grouped as a "chunk". Humans automatically adopt this strategy to save themselves time. For instance, random numbers had to be used instead of text in the example above, because people do not type their native language as individual characters. Instead, they combine the letters into words and remember these chunks instead. Put another way, the characters are no longer considered as individual choices.
A well designed interface should promote the use of chunking as a strategy by the user. One easy way is to gather together related options in a single place. This is one reason that like commands are grouped into a single menu which is hidden except for its title. If all of the menu options were "in the open", the user would be overwhelmed with dozens of alternatives at once. Instead, a "Show Info" command, for instance, becomes two chunks: pick File menu, then pick Show.
Sometimes the interface can accomplish the chunking for the user. Consider the difference between a slider bar in a GEM program, and a three digit entry field in a text mode application. Obviously, the GEM user has fewer decisions to make in order to set the associated variable.
While we are puttering around trying to speed up the keyboard, the mouse, and the screen, the user is actually trying to get some work done. We need to back off now, and look at the ways of thinking, or cognitive processes, that go into accomplishing the job.
The user's goal may be to enter and edit a letter, to retrieve information from a database, or simply draw a picture, but it probably has very little to do with programming. In fact, the Problem Space Principle says that the task can be described as a set of states of knowledge, a set of operators and associated constraints for changing the states, and the knowledge to choose the appropriate operator, which resides in the user's head.
Those with a background in systems theory can consider this as a somewhat abstract, but straightforward, statement in terms of state variables and operators. A programmer might compare the knowledge states to the values of variables, the operators to arithmetic and logic operations, the constraints to the rules of syntax, and the user's knowledge to the algorithm embodied by a program.
A rational person will try to attain his goals (get the job done) by changing the state of his problem space from its initial state to the goal state. The initial state, for instance, might be a blank word processor screen. The desired final state is to have a completed business letter on the screen.
The Rationality Principle says that the user's behaviour in typing, mousing, and so on, can be explained by considering the tasks required to achieve the goal, the operators available to carry out the tasks, and the limitations on the user's knowledge, observations, and processing capacity. This sounds like the typical user of a computer program must spend a good deal of time scratching his head and wondering what to do next. In fact, one of Card and Moran's key results is that this is NOT what takes place.
What happens, in fact, is that the trained user strikes a sort of "modus vivendi" with his tool and adopts a set of repetitive, trained behavior patterns as the best way to get the job done. He may go so far as to ignore some functions of the program in order to set up a reliable pattern. What we are looking for is a way of measuring and predicting the "quality" of this trained behaviour. Since using computers is a human endeavour, we should consider not only the speed with which the task is completed, but the degree of annoyance or pleasure associated with the process.
Card and Moran constructed a series of behavioural models which they called GOMS models, for Goals-Operators-Methods-Selection. These models suggested that in the training process the user learned to combine the basic operators in sequences (chunks!) which then became methods for reaching the goals. Then these first level methods might be combined again into second level methods, and so forth, as the learning progressed.
The GOMS models were tested in a lengthy series of trials at Xerox PARC using a variety of word processing software. (Among the subjects of these experiments were the inventors of the windowing methods used in GEM!) The results were again surprising: the level of detail in the models was really unimportant!
It turned out to be sufficient to merely count up the number of keystrokes, mouse movements, and thought intervals required by each task. After summing up all of the tasks, any extra time for the computer to respond, or the user to move his hands from keyboard to mouse, or eyes from screen to printed page is added in. This simplified version is called the Keystroke-Level Model.
As an example of the Keystroke Model, consider the task of changing a mistyped letter on the screen of a GEM word processor. This might be broken down as follows: 1) find the letter on the screen; 2) move hand to mouse; 3) point to letter; 4) click mouse button; 5) move hand to keyboard; 6) strike "Delete" key; 7) strike key for new character.
The sufficiency of the Keystroke Model is great news for our attempt to design faster interfaces. It says we can concentrate our efforts on minimizing the number of total actions to be taken, and making sure that each action is as fast as possible. We have already discussed some ways to speed up the mouse and keyboard actions, so let's now consider how to speed up the thought intervals, and cut the number of actions.
One way to cut down "think time" is to make sure that the capacity of short-term memory is not exceeded during the course of a task. For example, the fix-a-letter task described above required the user to remember 1) his place in the overall job of typing the document; 2) the task he is about to perform; 3) where the bad character appeared, and 4) what the new character was. When this total of items creeps toward seven, the user often loses his place and commits errors.
You can appreciate the ubiquity of this problem by considering how many times you have made mistakes nesting parentheses, or had to go back to count them, because too many things happened while typing the line to remember the nesting levels. The moral is that operations with long strings of operands should be avoided when designing an interface.
The single most important factor in making an interface comfortable to use is increasing its predictability, and decreasing the amount of indecision present at each step during a task. There is (inevitably) an Uncertainty Principle which relates the number of choices at each step to the associated time for thought:
T = I * LOG2 ( N + 1)
where LOG2 is the binary logarithm function, N is the number of equally probable choices, and I is a constant of approximately 140 msec/bit. When the alternates are not equally probable, the function is more complex:
T = I * SUM-FOR-i-FROM-1-TO-N (P(i) * LOG2( 1 / P(i) + 1) )
where the P(i) are the probabilities of each of the choices (which must sum to one). (SUM-FOR-i… is the best I can do for a sigma operator on-line!) Those of you with some information theory background will recognize this formula as the entropy of the decision; we'll come back to that later.
So what can we learn from this hash? It turns out, as we might expect, that we can decrease the decision time by making some of the user's choices more probable than others. We do that by means of feedback cues from the interface.
The important of reliable, continuous meaningful feedback cannot be emphasized enough. It helps the beginner learn the system, and its predictability makes the program comfortable for the expert. Programs with no feedback, or unreliable cues, produce confusion, dissonance, and frustration in the user.
This principle is so important that I going to give several examples from common GEM practice. The Desktop provides several instances. When an object is selected and a menu drops down, only those choices which are legal for the object are in black. The others are dimmed to grey, and are therefore removed from the decision. When a pick is made from the menu, the bar entry remains black until the operation is complete, reassuring the user that the correct choice was made. In both the Desktop and the RCS, items which are double-clicked open up with a "zoom box" from the object, again showing that the right object was picked.
Other techniques are useful when operator icons are exposed on the screen. When an object is picked, the legal operations might be outlined, or the bad choices might be dimmed. If the screen flashing produced by this is objectionable, the legal icons can be made mouse sensitive, so they will "light up" when the cursor passes over - again showing the user which choices are legal.
The desire for feedback is so strong that it should be provided even while the computer is doing an operation on its own. The hour glass mouse form is a primitive example of this. More sophisticated are "progress indicators" such as animated thermometer bars, clocks, or text displays of the processing steps. The ST Desktop provides examples in the Format and Disk Copy functions. The purpose of all of these is to reassure the user that the operation is progressing normally. Their lack can lead to amusing spectacles such as secretaries leaning over to hear if their disk drives are working!
Another commonly overlooked feature is error prevention and correction. Card and Moran's results showed that in order to go faster, people will tolerate error rates of up to 30% in their work. Any program which does not give a fast way to fix mistakes will be frustrating indeed!
The best way to cope with an error is to "make it didn't happen", to quote a common child's phrase. The same feedback methods discussed above are also effective in preventing the user from picking inappropriate combinations of objects and operations. Replacement of numeric type-ins with sliders or other visual controls eliminates the common "Range Error". The use of radio buttons prevents the user from picking incompatible options. When such techniques are used consistently, the beginner also gains confidence that he may explore the program without blundering into errors.
Once an error has occurred, the best solution is to have an "inverse operation" immediately available. For instance, the way to fix a bad character is to hit the backspace key. If a line is inadvertently deleted, there should be a way to restore it.
Sometimes the mechanics of providing true inverses are impractical, or end up cluttering the interface themselves. In these cases, a global "Undo" command should be provided to reverse the effect of the last operation, no matter what it was.
Now I am going to depart from the Card, Newell and Moran thread of discussion to consider how we can minimize the number of operations in a task by altering the modes of the interface. Although "no modes" has been a watchword of Macintosh developers, the term may need definition for Atarians.
Simply stated, a mode exists any time you cannot get to all of the capabilities of the program without taking some intermediate step. Familiar examples are old-style "menu-driven" programs, in which user must make selections from a number of nested menus in order to perform any operation. The options of any one menu are unavailable from the others.
Recall that the user is trying to accomplish work in his own problem space, by altering its states. A mode in the program adds additional states to the problem space, which he is forced to consider in order to get the job done. We might call an interface which is completely modeless "transparent", because it adds no states between the user and his work. One of the best examples of a transparent program is the 15-puzzle in the Macintosh desk accessory set. The problem space of rearranging the tiles is identical between the program and a physical puzzle.
Unfortunately, most programmers find themselves forced to put modes of some sort into their programs. These often arise due to technological limitations, such as memory space, screen "real estate", or performance limitations of peripherals. The question is how the modes can be made least offensive.
I will make the general claim that the frustration which a mode produces is directly proportional to the amount of the user's bandwidth which it consumes. In other words, we need to consider how many keystrokes, mouse clicks, eye movements, and so on, are going into manipulating the true problem states, and how many are being absorbed by the modes of the program. If the interface is wasting a large amount of the user's effort, it will be perceived as slow and annoying.
Here we can consider again the hierarchy of goals and methods which the user employs. When the mode is low in the hierarchy, and close to the user's "fingertips", it is encountered the most frequently. For instance, consider how frustrating it would be to have to hit a function key before typing in each character!
The "menu-driven" style of programs mentioned above are almost as bad, since usually only one piece of information is collected at each menu. Such a program becomes a labyrinth of states better suited to an adventure game!
The least offensive modes are found at the higher, goal related levels of the hierarchy. The better they align with changes in the state of the original problem, the more they are tolerated. For example, a word processing program might have one screen layout for program editing, another for writing letters, and yet another while printing the documents. A multi-function business package might have one set of menus for the spreadsheet, another for a graphing module, and a third for a database.
In some cases the problem solved by the program has convenient "fracture lines" which can be used to define the modes. An example in my own past is the RCS, where the editing of each type of resource tree forms its own mode, with each of the modes nested within the overall mode and problem of composing the entire resource tree.
Any narrative description of user interface is bound to be lacking. There is no way text can convey the vibrancy and tactile pleasure of a good interface, or the sullen boredom of a bad one. Therefore, I encourage you to experiment. Get out your favourite arcade game and see if you can spot some of the elements I have described. Dig into your slush pile for the most annoying program you have ever seen, run it and see if you can see mistakes. How would you fix them? Then… go do it to your own program!
This concludes the sermon. I'd like some Feedback as to whether you found this Boring Beyond Belief or Really Hot Stuff. If enough people are interested, homily number two will appear a few episodes from now. The very next instalment of ST PRO GEM will go back to basics to explore VDI drawing primitives. In the meantime, you might investigate some of the Good Books on interface design referenced below.
Stuart K. Card, Thomas P. Moran, and Allen Newell, THE PSYCHOLOGY OF HUMAN-COMPUTER INTERACTION, Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1983. (Fundamental and indispensible. The volume of experimental results make it weighty. The Good Parts are at the beginning and end.)
"Macintosh User Interface Guidelines", in INSIDE MACINTOSH, Apple Computer, Inc., 1984. (Yes, Atarians, we have something to learn here. Though not everything "translates", this is a fine piece of principled design work. Read and appreciate.)
James D. Foley, Victor L. Wallace, and Peggy Chan, "The Human Factors of Computer Graphics Interaction Techniques", IEEE Computer Graphics (CG & A), November 1984, pp. 13-48. (A good overview, including higher level topics which I have postponed to a later article. Excellent bibliography.)
J. D. Foley and A. Van Dam, FUNDAMENTALS OF INTERACTIVE COMPUTER GRAPHICS, Addison Wesley, 1984, Chapters 5 and 6. (If you can't get the article above, read this. If you are designing graphics apps, buy the whole book! Staggering bibliography.)
Ben Schneidermann, "Direct Manipulation: A Step Beyond Programming Languages", IEEE Computer, August 1983, pp. 57–69. (What do Pac-Man and VisiCalc have in common? Schneidermann's analysis is vital to creating hot interfaces.)