Classifying Human Control Behaviour by Artificial Intelligence

2019 
Traditional analysis of human manual control behaviour is currently constrained by the linear time-invariant assumption of the state-of-the-art “cybernetic” approach. This implies that time-varying and nonlinear aspects of human behaviour are generally overlooked, while these could be critical characteristic factors in, for example, the adaptation of human manual control behaviour due to variations in controlled element dynamics. Classification with Neural Networks would enable capturing both the nonlinear and time-varying human manual control behaviour characteristics, an approach that is absent in current literature. In this thesis, a Long Short Term Memory (LSTM) Recurrent Neural Network is used for classifying the adaptation of human manual control behaviour between single integrator (rate control) and double integrator (acceleration control) controlled element dynamics based on small time-samples of human control data. Datasets from two different previous human-in-the-loop experiments with compensatory manual tracking tasks are used to train the classification model and assess its capabilities. In addition, an application of the proposed model is tested on data from a third experiment, with a compensatory tracking task with time-varying controlled element dynamics. The LSTM model achieves 96% accuracy in classifying single/double integrator control behaviour, as well as for the application to detection of time-varying controlled element dynamics. This makes this approach a promising extension of the traditional “cybernetic” methods for (online) detection of adaptations in human manual control behaviour.
    • Correction
    • Source
    • Cite
    • Save
    • Machine Reading By IdeaReader
    0
    References
    0
    Citations
    NaN
    KQI
    []