Fine-tuning human-machine cooperation: Cognitive processing oriented assistance coordination.

摘要

With emergence of increasingly complex human-machine interaction systems, traditional machines are becoming more reliant on artificial intelligence. In human-machine cooperation, automated assistance promises to simultaneously reduce operators' workload and human errors. In fact, "intelligent" machines frequently encounter difficulties in complex dynamic environments due to limited human-like adaptability. Chaotic and unreliable assistance usually leads to cognitive overload, emotional suffering, and other negative behavior manifestations. Such negative consequences significantly magnify the importance of effective management of human-machine cooperation.;By analyzing human cognitive information processing characteristics, this thesis proposes a framework to unify the assistive measures that have very different manifestations on the cognitive basis. Therefore, human-machine cooperation is simplified to provide coordinated cognitive assistance to meet human operators' cognitive demands. This constitutes the most fundamental contribution of this research. Cognitive assistance is classified into indirect and direct approaches. Indirect cognitive assistance is comprised of operators' mental states management or adjustments -- such as visual/auditory attention adjustment, mental workload management, and emotion regulation. Direct cognitive assistance is classified into three levels: soft aid, soft intervention, and hard intervention, based on the assistive stages of human cognitive processing.;Three particular issues arising within the framework of cognitive assistance are further addressed in this study. Firstly, the downward U-shape relationship between emotional dimensions (arousal and valence) and human performance, based on an emotion-performance relation model for designing operators' emotion-aware machines, is proposed. Secondly, a novel approach of allowing intelligent machines to build humanlike confidence to convey limited reliability is proposed. The perceivable confidence expression helps operators to achieve better trust tuning and decision fusion. Lastly, when direct cognitive assistance is viewed as a two-dimension problem, a control approach is proposed to coordinate multi-source cognitive assistance in which operators' task performance and mental states are considered simultaneously.;This research conducted several driving assistance experiments to validate the above hypotheses and methods. Cognitive assistance exposed advantages for simultaneously improving human performance and maintaining a positive human-machine relationship by taking operators' mental states into account. These methods can be further generalized to manage operator assistance systems in many other industrial applications to improve their usability.