Modeling Human Control Behavior in Command-following Tasks

Humans interact with a variety of complex dynamic systems on a daily basis. However, they are often the lesser understood component of human-in-the-loop (HITL) systems. In this dissertation, we present the results of two HITL experiments to investigate the control strategies that humans use when performing command-following tasks. The first experiment is designed to investigate the control strategies that humans use to interact with nonlinear dynamic systems. Two groups of human subjects interact with a dynamic system and perform a command-following task. One group interacts with a linear time-invariant (LTI) dynamic system and the other group interacts with a Wiener system, which consists of the same LTI dynamics cascaded with a static output nonlinearity. In the second experiment, we examine the impacts of a relaxed command-following control objective on the control strategies used by humans. Two groups of human subjects interact with the same dynamic system and perform a command-following task; however, the groups have different control objectives. One group\u27s control objective is to follow the reference command as closely as possible at all times, while the other group\u27s control objective is to follow the reference command with some allowable error. We develop and utilize a new subsystem identification (SSID) algorithm to model control behavior of the human subjects participating in these HITL experiments. This SSID algorithm can identify the feedback and feedforward controllers used by human subjects, and is applicable to both linear and nonlinear dynamic systems. The SSID results of the first experiment indicate that adaptive feedforward inversion is the main control strategy used by human subjects for both linear and nonlinear plants. The results of the second experiment suggest that not all the human subjects who are instructed to perform a relaxed command-following task adopt adaptive feedforward inversion as their primary control strategy. The control behavior of those human subjects contains significant nonlinearities, which cannot be captured by a LTI control model. We present a nonlinear feedforward control architecture that can model several aspects of their control behavior