Identifying the time-varying, adapting human operator online could enable adaptive human-machine support systems and attention monitoring for human-in-the-loop vehicle control tasks. A validated, online identification method, including adaptation detection is missing, however.

In this MSc thesis project, online human operator identification using low-order ARX models with recursive least squares estimation was implemented. Operator delay estimation was attempted in simulations using the parallel recursive ARX method. Two online adaptation detection algorithms were evaluated on experimental data from eight subjects who performed a compensatory tracking task with time-varying controlled element. Participants were instructed to explicitly indicate when they detected this change with a button push, as comparison with the performance of the detection methods.

The experiment validated the online identification method, but time-varying delay estimation was not sufficiently accurate. The best detection method was based on recognizing the deviation of the human error rate response gain from a priori measured non-adapted behavior, and had a detection accuracy of 57%. A more practical method based on the parameter moving average was also developed and tested, but had a lower accuracy of 43%. For both methods, the lag in detection was found to be equivalent to the human operator detection lags measured in the experiment. Thus, the developed methods open the door for further development of identification-driven adaptive operator support systems in manual vehicle control tasks.