Understanding the developmental process of joint attention related actions, such as gaze following and alternation, is one of essential issues for the emergence of communication. Previous synthetic studies have proposed learning methods for gaze following without any explicit instructions as the first step to understand the development of these actions. However, a robot was given a priori knowledge about which pair of sensory information and action should be associated. This paper addresses the development of social actions without such knowledge with a learning mechanism that iteratively acquires social actions by finding and reproducing the contingency inherent in the interaction with a caregiver. The measurement of contingency based on transfer entropy is used to find appropriate pairs of variables for acquiring social actions from possible candidates. The reproduction of found contingency promotes a change of contingent structure in the subsequent actions of a caregiver and a robot. In computer simulations of human-robot interaction, we examine what kinds of actions related to joint attention can be acquired in which order by controlling the behavior of caregiver agents. The result shows that a robot acquires joint attention related actions in an order that resembles an infantpsilas development of joint attention.