Briefly, we have developed a comprehensive on line text-manipulation system. We wanted to determine the best means by which a user can designate textual entities to be used as "operands" in the different text-manipulation operations.

Techniques and devices for display-entity operand selcction represent a major component in any display control scheme, and are readily isolated for purposes of comparative testing, once the procedural environment in which selection is done has been established.

An important conclusion of our experimentation is that this environment has considerable effect upon the choice of display-selection means for a given display-control system.

To emphasize this, we point out that for two years we have been using the system for producing most of the internal memos* -- *and all of the proposals and reports -- associated with our research program.

This paper itself was extracted from one of these reports -- reorganized and modified by use of the system. See 1 (ENGLISH 1).

The format and writing style which represent an important experimental component of our research, are left in the form with which we work.

The user has generally been entering information on the typewriter-like keyboard.

To begin making the screen selection, his righl hand leaves the keyboard and takes hold of ("accesses," in our terminology) the selection device.

By moving this device he controls the position on the screen of an associated tracking mark (or "bug"), placing it over the "target" text entity.

He then actuates a pushbutton associated with the particular selection device, to tell the computer that he is now "pointing at" the target entity.

The computer puts a special mark under the entity which it determines as having been selected, to give the user an opportunity to see if a correct selection has been made.

The comparative speed with which they could be used to select material on the display screen. Two kinds of time period were measured:

The comparative ease with which an untrained user could become reasonably proficient in using the various devices.

The comparative error rates of the various devices.

Operand entities displayed on the screen are chosen by selecting a character within the operand entity (word, line, or statement).

The light pen or bug is first located near the desired character, then the SELECT switch on the device is depressed (or in the case of the knee control a special "CA" key on the keyboard is struck).

The Grafacon was manufactured by Data Equipment Company as a graphical input device for curve tracing (see Figure 2. See 2 (FLETCHER 1). The particular device that we tested is no longer marketed under this name. Data Equipment Company now markets the Rand Tablet under the name "Grafacon." See 3 (DAVIS 1).

It consists of an extensible arm connected to a linear potentiometer, with the housing for the linear potentiometer pivoted on an angular potentiometer.

A knob on the Grafacon arm is moved about by the user, and is depressed to activate the select switch (added by SRI) associated with the Grafacon.

The joystick that we used was manufactured by Bowmar Associates (Model X-2438; see Figure 2).

It is constructed from two potentiometers, mounted perpendicularly and coupled to a vertical stick in such a way that they resolve the motion of the stick into two components.

Two modes of operation with the joystick were implemented:

The "mouse" was developed by SRI in connection with this research.

It is constructed from two potentiometers, mounted orthogonally, each of which has a wheel attached to its shaft (see Figure 3).

As the case is moved over a surface (e.g., the table surface in front of a display)

A travel of about five inches is required for full edge-to-edge or top-to-bottom coverage of the CRT screen.

A switch mounted on the case is used for the select function.

A preliminary model of a knee control was made for this research (see Figure 4).

It consists of two potentiometers and associated linkage plus a knee lever. The linkage is spring-loaded to the right and gravity-loaded downward.

The user pushes the lever with his knee; a side to-side motion of the knee moves the bug edge-to-edge, while the top-to-bottom bug movement is controlled by an up-and-down motion of the knee (i.e., a rocking motion on the ball of the foot).

The light pen used was manufactured by Sanders Associates of Nashua, New Hampshire (Model EO-CH; see Figure 5).

It consists of a hand-held pen coupled to a photo multiplier tube by a fiber optic bundle.

The pen is pointed at the desired character on the CRT screen with the aid of a projected circle of orange light indicating the approximate field of view of the lens system.

However, certain features of the live working conditions were not closely related to the actual efficiency of the operand-selecting devices, such as

We tried either to eliminate these features from the experimental environment, or to fix them in some standard way through the experiment

A configuration simulating the "character mode" operation of the system consisted of nine x's, in a three by three array, with the array as a whole randomly placed on the display. The specific target entity was the middle x (see Figure 6(a)).

A configuration simulating the "word mode" operation of the system consisted of nine groups of five x's each, in a three by three "word" array, with the array as a whole randomly placed on the display. The target entity was any one of the five middle x's (i.e., any character in the middle " word"; see Figure 6(b)).

When the target appeared on the display screen, the subject was to strike the keyboard space-bar with his right hand, causing the bug to appear on the display. (Requiring that he use his right hand for both the space bar and the operand-selecting device made the experimental task closer to the actual on-line environment, where the user would often have both hands at the keyboard before moving to the operand selecting device. It also gave us a way of measuring the access times for the various devices.)

The subject was then to move his hand to the bug-positioning device being tested, and use it to guide the bug to the target entity on the display.

When the bug and the target coincided the subject was to "fix" the bug at that location, using the select switch of the bug-positioning device.

When the light pen rather than a bug-positioning device was used, the task sequence was much the same: after the target appeared, the subject was to strike the keyboard space bar with his right hand, then grasp the light pen and point it at the target entity (with the aid of the finder beam). The subject "fixed" his choice by depressing the select switch on the light pen. Correct and incorrect selections were signaled in the same way as with the bug-positioning devices.

For the experienced subjects, the entire testing procedure, which was broken into two time periods proceeded as follows:

For inexperienced subjects, the experimental procedure was somewhat different:

The position of the bug (in relation to the target entity) was recorded each 10 mulliseconds.

The times the subject hit the space bar, and the times he made either a correct or an incorrect entity selection, were recorded and appropriately tagged to aid in identifying these significant points in the late data analysis.

Tape-handling operations, controlled by commands from the on-line keyboard, facilitate searching through the data recorded on the magnetic tapes. These commands allowed one to scan forward or backward by one 32-target block of tests (or, an 8-target block, in the records for inexperienced subjects); and, within that block, to scan forward or backward one target (i.e. one presentation-selection event) at a time.

For each target-fix, the CRT could display a graph showing the bug's distance from its target entity as a function of time. This was displayed as two curves (see Figure 7), one showing variation with time of horizontal distance, and the other of vertical distance. The time-count was begun when the target appeared on the display. Vertical lines on the curves mark the time at which the space bar was struck and the time at which the target was correctly selected. Incorrect selections are shown as x's on the curve.

When studying a given target-fix event, the experimenter can, if he wishes, initiate output (to the on-line typewriter) of performance data: the time at which the space bar was struck, the time at which the bug movement began, the time at which the target was correctly selected, and the number of errors (incorrect selections) made. This software also computed and printed out the following incre mental times: the access time (from the time the space bar was struck until the time the bug movement began, measuring how long it took the subject to move his hand from the keyboard to the device); the motion time (from the time the bug began moving until the time the target was correctly selected); and total time (from the time the space bar was struck until the time the target was correctly selected -- i.e., the sum of access time plus motion time).

Finally, there is another command which causes the computer to search through a 32-target block of target fixes and compute (for output to the on-line typewriter) the average incremental times, and total number of errors, for that block.

Figures 8 and 9 are taken from the results of the eight experienced subjects, some of whom were very familiar with the on-line system and had used the devices often. Figure 8(a) shows the average total time (for all experienced subjects) required for a correct selection of the "character" target, with no penalty for errors; Figure 8(b) shows the results of the same tests with a 30 percent penalty for errors. Figure 9(a) and Figure 9(b), respectively, show the same for the "word" target.

Figure 10 shows the results from the tests of subjects who had had no previous experience with the devices. Figure 10(a) imposes no penalty for errors. Figure 10(b) imposes a 30 percent penalty for errors, as explained above.

Figures Figure 11 and Figure 12 compare the error rates for the various devices. Figure 11 shows the results for the "character" and "word" tests, as performed by experienced subjects (using four different operand locating devices); Figure 12(a) shows the results of the "character" tests for inexperienced subjects (using six different operand-locating devices).

As mentioned above, the knee control was not developed soon enough to include it in the tests for the experienced subjects (where we included only devices that had been available for some time, in order to avoid bias). We did, however, perform a few individual check tests with experienced subjects, using the knee control; in these tests the knee control appeared both slower and less accurate than the light pen and mouse.

Inexperienced subjects found the knee control was the fastest device. Undoubtedly the main reason for this was that the knee control, unlike all the others, has no access time. (If the access time is subtracted from the total times measured for the other devices, the knee control no longer show up so favorably.)

Inexperienced subiects also found the light pen faster than the mouse. A reason for this may be that the light pen exploits one's inherent tendency to select something by straightforwardly "pointing" at it rather than by guiding a bug across a screen toward it from a remote control. This means that an inexperienced subject can become reasonably proficient in using a light pen with relatively little practice.

The joystick proved to be both the slowest and the least accurate of the devices we tested, in both modes of its operation ("absolute" and "rate"), and among both the experienced and inexperienced subjects.

It is interesting to note, however, that both the joystick and the Grafacon showed up more favorably (relative to the other devices) when used to select word entities rather than character entities. These two devices seem to perform better where fine control is less critical; they can move into range quickly at the grosser level.

Both the Grafacon and joystick suffer from a lack of independence in the actions required to actuate the select switch and to move the bug. By contrast, the mouse is moved by an action of the entire hand, while the switch is easily operated by one finger and does not tend to cause bug motion.

With the joystick the scale factor between bug motion and device motion was about 4:1 for a normal finger position on the stick; for the mouse and Grafacon, the scale was about 2:1. This may have contributed to the lack of fine control (and high error rate) for the joystick.

The rate mode with the joystick is very poor, partly because of the software implementation.

The light pen may have showed up poorly for several reasons.

However, the results compiled and plotted to test this hypothesis did not show any significant correlation.

An examination of the CRT-displayed perform ance curves suggests that this may be because the time to move the bug close to the target is relatively small compared to the average access time, and to the average time required for selecting the target after the bug has been moved close to it.

In using the Grafacon and the joystick (rate mode), the subjects tended to overshoot the target, losing a significant amount of time in changing the bug's direction and bringing it back into position for a select action.

While our experiments did not provide a measure of access time for the light pen, we found (from observing the subjects) that a good deal of time was consumed in reaching from the keyboard to grasp the light pen.

Though the knee control showed up well in its performance as compared with the other devices, an examination of its CRT-displayed curves shows that its operation is relatively unsmooth; the bug tends to move erratically, and it appears to be difficult to move the bug vertically on the display.

The operand-selecting devices that showed up well in our tests were the mouse; the knee control; and the light pen. These three were generally both faster and more accurate than the other devices tested.

Inexperienced subjects did not perform quite as well with the mouse as with the light pen and knee control, but experienced subjects found the mouse the "best" of the devices tested, and both groups of subjects found that it was satisfying to use and caused little fatigue.

The select switches on both the Grafacon and joystick tended to move the bug and cause an incorrect fix. These two devices could probably be improved by redesigning their select switch mechanisms.

Although the knee control was only primitively developed at the time it was tested, it ranked high in both speed and accuracy, and seems very promising. It offers the major advantage that it leaves both hands free to work at the keyboard.

The major advantage of the light pen appeared to be its psychological "naturalness" of operation in pointing at the item to be selected. This means that an untrained user can quickly understand it and gain enough proficiency to do useful work.

The principal value of our experimental work to date was in developing the techniques of experiment and analysis, and in isolating some of the factors in the design of display-selection means that are important to fast operation.

Any comparative evaluation of the different types of devices must be qualified to such an extent that it is not significantly useful in a direct sense toward choosing from among the types of devices.

What is important to fast, efficient display selection is the particular feel to the user of the thing he grasps and moves, e.g.:

To make final judgments between display-selection devices, more must be learned about the desirable way to adjust and coordinate each of these factors. Then it must be seen which basic-device approach can best provide this.

Comparative comments of a general sort can follow these observations:

An important, general conclusion from our tests is that the relative value of different schemes cannot be judged on the basis of their appeal to inexperienced users.

This "lesson" can be expressed as follows:

Thus, it seems unrealistic to expect a flat statement that one device is better than another. The details of the usage system in which the device is to be embedded make too much difference.