HCI International 2005 - The future of augmentation managers

In every Augmented Cognition (AugCog) application, where a closed-loop architecture is implemented, there is a requirement for a system that is responsible for the implementation of the strategies designed to mitigate the effects of excessive workload. The manner by which these mitigation strategies are implemented is likely to be platform specific, however several generic principles may be defined which enable the outputs of operator state gauges to be interpreted and acted upon.

In the simplest configuration an Augmentation Manager is designed to implement a single or number of mitigation strategies based on a single, high-level descriptor of operator state. In the initial implementation of the Tasking Interface Manager (TIM) in our Cognitive Cockpit (CogPit) the output of a single gauge, executive workload, was passed through a linear filter and a simple thresholding algorithm employed to trigger the implementation of a number of mitigation strategies. Whilst this showed some benefits during our first set of closed-loop trials, it is likely that this approach oversimplifies the problem. In particular, this approach did not take account of the complex relationship between mitigation strategies and state. In addition, any system that employs this, or similar techniques, will be purely reactive rather than proactive. Here we propose a multifaceted approach, in which the outputs of multiple gauges, operator state forecasting and context modeling together with a prediction of the effects of mitigation strategies on operator status will be used to drive the targeted application of a number of mitigation strategies. Achieving this will require a major engineering effort and carefully constructed empirical research.

Furthermore it is apparent that in order to make AugCog systems truly intelligent and adaptive, they will need to be able to monitor their own performance and adapt as necessary. Future systems will also need to be adaptive to the context in which they are operating, enabling the appropriate responses to unfamiliar contextual cues to be made. These occurrences should be stored in order to improve system operation under similar conditions should they arise again. This machine learning element to AugCog systems is probably the most underdeveloped area, but potentially could provide huge benefits to their long-term operational functionality.