Objective Model Selection for Identifying the Human Feedforward Response in Manual Control

Abstract

Abstract-Realistic manual control tasks typically involve predictable target signals and random disturbances. The human controller (HC) is hypothesized to use a feedforward control strategy for target-following, in addition to feedback control for disturbance-rejection. Little is known about human feedforward control, partly because common system identification methods have difficulty in identifying whether, and (if so) how, the HC applies a feedforward strategy. In this paper an identification procedure is presented that aims at an objective model selection for identifying the human feedforward response, using linear time-invariant ARX models. A new model selection criterion is proposed to decide on the model order (number of parameters) and the presence of feedforward in addition to feedback. For a range of typical control tasks, it is shown by means of Monte Carlo computer simulations that the classical Bayesian Information Criterion (BIC) leads to selecting models that contain a feedforward path from data generated by a pure feedback model: 'false-positive' feedforward detection. To eliminate these false-positives, the modified BIC includes an additional penalty on model complexity. The appropriate weighting is found through computer simulations with a hypothesized HC model prior to performing a tracking experiment. Experimental human-in-theloop data will be considered in future work. With appropriate weighting, the method correctly identifies the HC dynamics in a wide range of control tasks, without false-positive results.

Effects of Preview Time in Manual Tracking Tasks

“…Longer preview times may trigger the near-viewpoint response, as the HC's knowledge of the target signal improves. With 1 s of preview, most subjects were found to apply a near-viewpoint response in SI tasks, but not in DI tasks [14], as accurate feedforward control with these unstable dynamics is difficult [24], [25]. Although not all subtleties of the near-viewpoint response are yet understood, its dynamics appear to approximate a simple differentiator [14]:…”

Effects of Target Signal Shape and System Dynamics on Feedforward in Manual Control

“…Note that averaging might blur distinctive features of feedforward, see [14]. Here, we insist on averaging over five runs to be consistent with the ARX method [18].…”

“…2) Black-box ARX Identification: The control behavior during the ramp and parabola conditions are identified by means of the black-box ARX identification method of [18]. It was developed to objectively identify if and how the HC utilizes a feedforward and/or feedback control strategy, without making any prior assumptions concerning the HC dynamics.…”

“…Recent advances in system identification [18] now allow for the investigation of feedforward control in realistic control tasks. To investigate the adaptation of manual feedforward control behavior to the target waveform shape and the system dynamics, we consider two realistic target waveform shapes and three classes of vehicle-like system dynamics.…”

“…Two complementary system identification and parameter estimation methods are used to analyze experimental human-in-the-loop data. A recently developed blackbox ARX identification procedure, which is robust to falsepositive feedforward identification, is used to identify HC dynamics [18]. A time-domain parameter estimation method [19], using HC models based on the ARX results, is used to obtain further insight in the correct model structure and to observe changes in control behavior.…”