Predicting Information Needs: Adaptive Display in Dynamic Environments

摘要: Predicting Information Needs: Adaptive Display in Dynamic Environments Bradley C. Love (brad love@mail.utexas.edu) Matt Jones (mcj@colorado.edu) Department of Psychology Austin, TX 78712 USA Boulder, CO 80309 Marc T. Tomlinson (mtomlinson@love.psy.utexas.edu) Michael Howe (michael.howe@mail.utexas.edu) Abstract canny ability to deliver information his commander mo- ments before the formulated request, much like how RADAR learns anticipate needs user reduce cognitive load. Before presenting and empirically evaluating it a well-controlled ex- periment, we briefly review related work. Although available human operators can in- crease without obvious bound, processing capacities remain fixed. Finding selecting relevant formation display this deluge options imposes burden on user. We describe domain-general system, Responsive Anticipates Requests (RADAR), that highlight would select if searched through all possible options. By offloading se- lection process RADAR, concentrate pri- mary task. Tests with subjects tank video game en- vironment required monitoring several chan- nels while maintaining situation awareness revealed play- ers performed better which channel display. customize its predictions take into account individual differences changes within over time. RADAR’s emphasis learning by observing minimizes need for explicit guidance from subject matter experts. Related Efforts The topic plan recognition AI is concerned cor- rectly attributing intentions, beliefs, goals Plan models tend subscribe Belief- Desires-Intention framework (McTear, 1993). This line work relies knowledge-based approaches mod- eling encoding insights domain-specific experts (Goodman & Litman, 1992). These involve identifying user’s subgoals task-analysis (Yi Bal- lard, 2006). Once are understood, be adapted appropriately Instead focusing internal state user, other rely input domain adapt emphasize should attend. For example label episodes these serve as training instances machine prioritize elements (St. John, Smallman, Manes, 2005). Alternatively, hu- man used build expert systems or Bayesian (Horvitz Barry, 1995). Our approach diverges aforementioned Rather than prescribe source prioritize, attempt Unlike recognition, sidestep problem ascribing ascertaining mental state. In- stead, directly predict desired dis- play contextual (i.e., situational) features. emphasizes opposed preprogrammed interfaces (M¨antyj¨arvi Sepp¨anen, 2002). Adopting ap- proach adaptive has number positive conse- quences, including individ- ual across users (Schneider-Hufschmidt, K¨uhme, Malinowski, Another consequence minimal system. Like keyhole (Albrecht, Zuker- man, Nicholson, 1998), our based observ- Introduction increasingly find ourselves information-rich environ- ments. Often, many sources potentially use- ful completing example, coordinating aster relief, useful include feeds, weather forecasts, inventories relief supplies, GPS tracking support vehicles, etc. Likewise, sensors, gauges, navigation modern auto- mobile driver. One key challenge people face at current moment. operator increase with- out basic capaci- ties Each additional incurs cost increasing complexity selection process. As informational channels added, some point, marginal costs (in terms load) eclipse benefits. In report, propose evaluate system eases highlighting nel approxi- mate ing behavior. cases where successfully approximates human’s process, offloaded RADAR. named after character Radar O’Reilly television series M*A*S*H. had an un-