Abstract: Cognitive load (CL) often becomes too great in modern human-machine systems, which affects the performance and reliability of system operators. Therefore an effective and reliable approach to measure CL is crucial. Currently there are three major measures of CL: these are subjective evaluation, performance measures and physiological assessment. None of these, however, can comprehensively reflect the state of CL under different task conditions. Thus, this study attempted to establish an appropriate multi-dimensional assessment technique to evaluate CL synthetically in the human-machine interaction process.
A dual-task experiment with simulated web-searching as the primary task and mental arithmetic as the secondary task was conducted. The primary task had three levels of complexity, high, medium, and low. Thirty-two subjects were selected to perform the dual -task with each complexity level of the primary task. The order of performance of the three complexity levels was randomized across subjects. Four types of indices were obtained, including primary-task measures, secondary-task measures, subjective evaluations and eye movement measures. The synthetic assessment models were constructed via factors analysis, BP artificial neural network and self-organizing feature map, respectively. The results showed that: (1) The six indices, which are mental effort, perceived task difficulty, duration of fixation, number of fixations, response time and the percentage of correct responses in the primary task, were sensitive to the change in cognitive load. (2) In dual-task situations, the multidimensional synthetic assessment of cognitive load generated more valid results than those based on the single assessment index. (3) According to the indices of relative, absolute and average errors, the two ANN models (BP neural network and self-organizing feature map) showed a more accurate measurement of cognitive load than the traditional technique of factors analysis

Key words: Human-machine interaction, cognitive load, assessment, neural network, modeling