Abstract: Complex operations required in modern human-machine systems often increase cognitive load on human operators. Under excessive cognitive load, human operators may exhibit delayed information processing or even make some mistakes, probably inducing various operating accidents. Therefore, an assessment and prediction of the general situation of the cognitive load will be needed to help design the system with proper cognitive load on the operators. Also, this will be of great significance to improve operators’ working environment, job satisfaction and operational security. The present study attempted to establish effective prediction models to evaluate cognitive load changes in human-machine interaction processes.
A dual-task experiment with simulated web-searching as the primary task and mental arithmetic as the secondary task was conducted. Twenty six volunteers, aged 20 to 27 years, participated in this study. Each participant performed sixteen phases of the task. The order of task phases was randomized across participants. Finally, three types of indices were obtained, including task-performance measures (response time and accuracy of the primary task), subjective evaluations (mental efforts and perceived task difficulty) and eye movement measures (gaze durations and the number of gazes). The effects of task duration on cognitive load were examined using these different indices. The BP and Elman neural network models were developed to predict the changing tendency of cognitive load in the different phases of the task.
The results showed that (1) Task duration had a significant effect on the response time and accuracy of the primary task, gaze durations, and the number of gazes respectively. However, it had no significant effect on either mental efforts or perceived task difficulty. (2) The BP and Elman neural network models could predict the changes in the relevant assessment indices of cognitive load in the different phases of the task. Furthermore, a synthetic analysis could be made in combination with the synthetic assessment model of cognitive load in the different phases of the task.
The current findings suggest that neural network models can be used to predict the cognitive load changes on human operators in complex human-machine systems efficiently. The BP and Elman neural network models developed here may be the good approaches to the dynamic self-adaptive allocation of tasks in modern complex systems, while the models developed in the current study may not be generally appropriate for different human-machine interactions. In a sense, the importance of the present findings is that they provide an effective approach as to how to predict the cognitive load changes in human-machine interaction processes.

Key words: human-machine interaction, prediction, cognitive load, neural network, modeling