Instructional efficacy of automated Conversational Agents designed to help small groups of students achieve higher learning outcomes can be improved by the use of social interaction strategies. These strategies help the tutor agent manage the attention of the students while delivering useful instructional content. Two technical challenges involving the use of social interaction strategies include determining the appropriate policy for triggering these strategies and regulating the amount of social behavior performed by the tutor. In this paper, a comparison of six different triggering policies is presented. We find that a triggering policy learnt from human behavior in combination with a filter that keeps the amount of social behavior comparable to that performed by human tutors offers the most effective solution to the these challenges.