Improved automatic adaptation through the combination of multiple information sources.

Abstract

Systems that automatically adapt to the needs of their human operators offer the potential to improve human-machine interactions in many applications. This research explores the use of a probabilistic reasoning construct known as Bayesian networks for combining multiple sources of information to improve adaptation decision accuracy. The first experiment employed a Bayesian network combining two information sources to distinguish between two types of tracking behavior in a manual tracking task. The Bayesian network made more accurate identifications than either information source did alone. A simulation demonstrated how the Bayesian network's performance was affected by the relative performance and reliability of the information sources. The testbed for the final two experiments was the NavChair Assistive Wheelchair Navigation System. The NavChair has three operating modes, each of which is appropriate for specific tasks. In the first experiment the NavChair was controlled by joystick and in the second by voice commands. In both experiments, the performance of the NavChair using a Bayesian network to combine location information and environmental cues to make operating mode selections (adaptation decisions) was compared to the NavChair's performance when: (1) environmental cues alone were used to make adaptation decisions, (2) the experimenter made adaptation decisions, and (3) no navigation assistance was offered to the wheelchair operator. The Bayesian network made more accurate adaptation decisions than environmental cues alone and was close to the decision accuracy of the experimenter making optimal adaptation decisions. Whether the performance of the NavChair with adaptive navigation assistance compared favorably with no navigation assistance depended on which control method was used. When subjects used a joystick, no navigation assistance was superior, as expected. However, when voice control was used (and performance degraded relative to that of joystick control), adaptive navigation assistance allowed better performance than no assistance. This research demonstrated that combining two information sources using probabilistic techniques can lead to more accurate adaptation decisions. However, improvement is not guaranteed but rather depends upon the interaction of the information sources. Future work is planned to further develop the NavChair and to apply the approach developed during this research to other adaptive human-machine systems.