Partial plan recognition from human-computer interactions under incomplete information

Date of Completion

January 1999


Computer Science




Networked computing has drastically changed the way in which people work and exchange information. In the evolution of the relationship between human and intelligent machines with the explosive network information available, the need of automated methods of gathering information and the need for higher-bandwidth interfaces by learning about and adapting to users are emerged. In our work, intelligent interface agents are developed for the prediction of resource usages in the UNIX domain, that is, assessing the likelihood of upcoming demands by users on limited resources by learning user behavior. ^ A multi-strategy approach is used for learning user regularities and making predictions on the regularities by integrating features of action reasoning and using mathematical user models based on hidden Markov models. The knowledge acquisition of user behavior is done by extracting behavioral patterns from their historical data of human-computer interactions. Predictive patterns are discovered by analyzing the correlations of actions and learned through segmentation and labeling of a sequence of actions. The issues of ambiguity, distraction and interleaved execution of user behavior are examined and taken into account to improve the probability estimation in hidden Markov models. ^ The work is a good example of bridging theory and practice: a formal model underneath, with development and verification based on real observations. Algorithms are developed for learning user regularities and the formal models are established in order to represent and solve the prediction problem on a theory basis. The prototype system “NOSTRODAMUS” is developed along with the design of algorithms and formal models. ^