Human skill evaluation by operation input characteristics with multiple neural networks

This paper addresses a human skill evaluation technique with characteristics of input command decision during operation tasks of a mobile object. The characteristics is how long past of tracking error information is utilized to decide the operation command. This evaluation is achieved by a novel human model with multiple neural networks which have different time series of input signals. Then, by the prediction errors from the neural networks, which time series is dominant to decide operation command can be estimated. In this paper, target tracking tasks in 2D CG environment are experimented. A main goal of the task is to operate the mobile target by joystick operation to keep approaching a reference target behaving randomly. Furthermore, effects of subliminal input filtering by predicted operation input are discussed. Finally, by analysis for distribution the neural networks prediction errors, the operator's skill is evaluated and verify the technique with conventional average of tracking errors during the task.