The communication rate of a person using a traditional single-switch scanning interface is highly dependent upon the inter-group and inter-item delays. We have developed a method for the automatic, real-time adjustment of scanning delays. Based upon quantitative measures of scanning performance such as the frequency of selection errors, the frequency of missed selections, and the portion of the delay utilized for selections, this adjustment scheme rapidly converges on an optimal set of scanning delays. The effectiveness of the technique was verified in a series of experiments. BACKGROUND The adjustment of the delays used for single-switch scanning interfaces has traditionally been an ad hoc procedure. Users or clinicians adjust the scanning delays based on approximate performance measures and the user's comfort level. In the course of an extended study on the effects of various scanning variations (e.g., with or without character prediction) on communication rate, we developed a method for dynamically adjusting the scanning delays based on quantitative aspects of user performance. These methods were integrated into our research software suite. Researchers have suggested various approaches to the optimization of scanning delays, but most can be distilled to a single rule: Decrease delays until the frequency of selection errors becomes unacceptable [1, 2]. For example, in one experiment the scientists stipulated that subjects make fewer than 5% incorrect selections and pass over correct selections less than 10% of the time, adjusting delays manually so as to satisfy these criteria [1]. Adopting a broader set of performance measures, we have developed a method of adjusting scanning delays without human intervention. APPROACH We have identified a core set of quantitative measures that indicate the appropriateness of a given scanning delay. These include selection errors, missed selections, missed prediction elements, and selection timing (as described individually below). The augmentative device automatically collects measures in each of these areas over a window of time that is long enough to provide a significant sampling of performance. The system then analyzes this information to determine whether the scanning delays should be increased, decreased, or left unchanged. We have adopted a proportional delay adjustment scheme in which the delays are multiplied by an adjustment factor whenever a change is required. When the parameters indicate that the delays should be shortened, they are multiplied by an adjustment factor less than one (in our studies, 0.95). Conversely, when the delays are to be increased a factor greater than one is applied (e.g., 1.05). Selection Errors: The most obvious indication that scanning delays are too short is frequent selection of undesired scanning groups or individual elements. Detecting group selection errors is fairly straightforward. Since most scanning interfaces provide a means to "unselect" groups that are inadvertently selected, the system need only monitor these unselection events. Detecting undesired element selections can be more problematic. We adopted the simple strategy of monitoring the use of isolated backspace or undo selections. Although corrections of this sort do not always indicate a Automatic Update of Scanning Delays scanning mistake, we found that non-repeated instances of corrections were highly predictive of selection errors. At the expense of system complexity, more reliable indicators could be devised. For example, one could examine the selection that occurred immediately after the correction to ensure that it was close (e.g., next in the scanning procession) to the original (incorrect) selection. Missed Selections: For most scanning interfaces, if a user misses a selection opportunity, he or she must wait for the scanning focus (cursor) to cycle back around. Missed selections of this sort are especially prevalent for the first one or two scanning groups (i.e., the first couple of rows and the first couple of columns). An excessive number of repeated scanning cycles in a given temporal window generally indicates that the scanning delay should be increased. With many devices, scanning will continue even when the user is not attending to the interface. Although repeated cycles of this type are not informative, positive identification of these situations is difficult. One ad hoc solution is to not count scanning cycles that repeat more than twice. Missed Prediction Elements : In some of our experimental interfaces, the static character matrix was supplemented by a dynamic character prediction list. This list was scanned character-bycharacter before proceeding to a row-column scan of the static matrix – an arrangement shown to produce the greatest switch savings [3]. Because the prediction matrix was dynamic, many subjects found that they needed longer delays for the predicted characters than they did for the other characters. If the delays were too short, the scanning focus passed over the desired character before the subjects had time to search for it the prediction list. When this happened the subjects selected the desired character from the static matrix – they missed the prediction element. A significant frequency of missed prediction elements indicates that the delays used for the dynamic portion of the scanning interface are too short. In the absence of other adjustment indicators (e.g., selection errors or missed selections), only the delays associated with the prediction list need be adjusted. As a practical consideration, however, users may prefer that all delays be adjusted for interface consistency. We considered only missed character predictions in our study, but this adjustment method can extended to handle missed words from word prediction lists. Selection Timing : The previous adjustment strategies are meant to compensate for various types of selection errors. If these errors occur frequently, the delay should be automatically increased. Conversely, if these errors occur infrequently it may be appropriate to decrease the scanning delays. We have also investigated an alternative strategy for decreasing delays based upon the fraction of the delay utilized during selection. If scanning is too slow for a user, individual selections will occur in the first part of the scanning delay. By analyzing the average fraction of the delay "used up" for each correct selection, we can determine if the delay can be shortened without jeopardizing selection accuracy. If the scanning delay is 2 seconds, but the user is consistently selecting elements only 500 milliseconds after they are first highlighted (25% of the delay), the delay can clearly be reduced. In practice, we found that delays could reliably be reduced without introducing selection errors if the average selection time was less than 65% of the delay. METHOD All of the automatic adjustment techniques described above were utilized in an experiment in which 15 able-bodied individuals each used two different single-switch scanning interfaces for 15 hours per interface [3]. Subjects were asked to transcribe a series of articles and to respond freely to a collection of non-intrusive written questions. Although the primary focus of the study was the effect of the scanning interface on communication rate, the performance of the dynamic adjustment scheme was an important area of secondary research. Automatic Update of Scanning Delays Scanning delays were initially set at 2 seconds. Delay adjustments were based on a window of 20 selections (including corrections). Scanning delays were multiplied by a factor of 1.05 when any one or more of the following conditions were met within the 20 selection window: (1) 3 or more element selection errors (isolated backspaces), (2) 3 or more unselected groups, (3) 3 or more separate repeated scanning cycles, (4) 3 or more missed prediction elements (where applicable). Scanning delays were adjusted by a factor of 0.95 when: (1) none of the preceding criteria were satisfied and (2) the average selection fraction fell below 0.65 (65%). RESULTS This plot depicts the average scanning delay as a function of the number of selections for a representative subject. Starting at 2 seconds, the delay drops off quickly, eventually stabilizing at 270 milliseconds. The plot is nearmonotonic until the final phase, at which point some mild oscillations occur – the system tries to decrease the scanning delay past the subject's capabilities, only to have it bumped up again in subsequent windows. Although subjects reported problems sustaining fast scanning rates, they did not report noticing the 25 to 50 millisecond oscillations. DISCUSSION We have established the effectiven ess of a series of dynamic adjustment procedures for scanning delays. Although these techniques were tested on character-based scanning arrays, most should apply equally well to word-based and symbol-based interfaces. Since the proposed adjustment methods rely upon device-collected performance measures (e.g., number of missed selections during a window of 20 selections), most cannot easily be incorporated into current hardware or software communication aids. Given the importance of optimizing scanning delays, however, we hope that manufacturers will begin to adopt these techniques. REFERENCES 1. Koester, H.H. & Levine, S.P. (1994). Learning and performance of able-bodied individuals using scanning systems with and without word prediction. Assistive Technology, 6, 42–53. 2. Szeto, A.Y.J., Allen, E.A., & Littrell, M.C. (1993). Comparison of speed and accuracy for selected electronic communication devices and input methods. Augmentative and Alternative Communication, 9, 229-242. 3. Lesher, G.W., Moulton, B.J., & Higginbotham, D.J. (1998). Techniques for augmenting scanning communication. Augmentative and Alternative Communication, 14, 81-101. ACKNOWLEDGMENTS This study was supported in part by a National Institute of Child Heal