Integrating probabilistic and knowledge-based systems for explanation generation uri icon


  • An important requirement for intelligent assistants is to have an explanation generation mechanism, so that the trainee has a better understanding of the recommended actions and can generalize them to similar situations. In this work we combine different knowledge sources to generate explanations for operator training. The explanations are based on a general template which is composed of 3 main parts: (i) the recommended action in the current situation; (ii) a graphical representation of the process highlighting the relevant variable; (iii) a verbal explanation. The optimal action is obtained from a Markov decision process (MDP) that guides the operator in the training session. For determining the relevant variable we developed a method that estimates the impact of each variable on the utility, selecting the one with highest impact in the current state. The verbal explanation is extracted from a domain knowledge base that includes the main components, actions and variables in the process, represented as a frame system. We present preliminary results of explanations generated in the power plant domain.

Publication date

  • December 1, 2008