Page 45 - Innovations in Intelligent Machines

Predicting Operator Capacity for Supervisory Control of Multiple UAVs 33
12 Fan-out (1)
Operator Capacity 8 Revised Fan-out (3)
10

6
4
2
0
1 2 3 4 5 6 7 8 9 10 11
Possible Number of UAVs
Fig. 14. Predictions Using Cost-Based Simulation Inputs

using the revised fan-out. Interestingly this number is very close to what was
experimentally observed in the previously described experiment.

4 Meta-Analysis of the Experimental and Modeling
Prediction methods

Two methods for determining maximum operator capacity for supervisory
control of multiple UAVs have been presented, both based on operator inter-
actions and wait times for mission tasks, as well as neglect times during which
one or more vehicles operate autonomously. The strengths and weaknesses of
each method will now be discussed, as well as how these methods could be
used synergistically.
In the first method, the original fan-out equation that related operator
interaction and vehicle neglect times (1) was revised to include operator wait
times (3). An experiment was conducted to determine if the revised fan-out
predictions more closely matched actual human-in-the-loop control scenarios.
The results showed that the revised fan-out model produced more conservative
estimates when modified to include wait times caused by human interactions,
which include interaction wait time, wait time in the queue, and wait time
due to the loss of situation awareness.
While this temporal-based method for computing fan-out gives more con-
servative general estimates, it lacks the cost-benefit analysis trade space rep-
resentation that can be found through optimization methods that provide for
sensitivity analysis. For example, in the experiment, it was estimated that
operators could control 7–16 UAVs in a low workload scenario, but only 3–7
vehicles in high workload settings. The ranges resulted from increasing levels
of automation as an experimental independent variable. Because these pre-
dictions were based on experimental data (which were discrete across four
different levels of automation), there can be no post-hoc sensitivity analysis,