When the human portion of RMSE in partitioned mode was compared to RMSE for the same axis in manual mode, a performance advantage of partitioned mode was not evident. Indeed, overall performance in partitioned mode was worse than in manual mode. However, an advantage of being paired with an expert- level computer was evident in the partitioned mode, when the human and computer were sharing the task. When participants fully controlled the task (manual mode), there were no performance differences between skill conditions.
The analysis of mean time between task mode allocations showed there was significantly less time spent in the partitioned mode overall than in the fully automatic or fully manual conditions. This result was unexpected and may be indicative of an overly narrow range around the mean engagement ("0.2 SD) used for partitioned mode. Figure 1 illustrates how the range around the mean engagement is related to system behavior. Any differences among the task modes, particularly in negative feedback, should show more mean time spent in partitioned mode, simply because the biocybernetic system attempts to maintain operator engagement close to the baseline mean. Recently, Scallen, Hancock, and Duley ( 1995) tested the effects of automation cycles on performance, and suggested imposing a damping factor, or governor, on an adaptive system which sets a minimum frequency with which changes in automation may occur. For the system used in the present study, a wider range around the baseline mean may be needed in order to increase the time between allocations. A follow-up to this study is addressing that issue.
The present study indicates a definite performance benefit for having a highly-skilled computer as a teammate. Thus, adaptive systems utilizing performance modeling should exhibit expert-level behavior to best aid the human operator.
Although the effects of task partitioning were examined in the present study, the extremely short periods that the task remained in partitioned mode prevented a valid assessment of this. A corrective measure may include widening the range around the baseline mean to limit switching frequency. A follow- up study is addressing that issue.
I would like to acknowledge my advisor, Dr. Mark Scerbo, and also Dr. Ray Comstock and Dr. Alan Pope at NASA Langley Research Center for their help with this research.
-152-