Our goal is to create a user-adaptive communication-robot. We are developing a system for evaluating human-robot interactions. Although such evaluation is indispensable for learning algorithms, users' preferences are too difficult to model because they are subjective. In this study, we used the interactive evolutionary computation (IEC) to configure the value system of a learning communicationrobot. The IEC is a genetic algorithm whose fitness function is performed by the user. In our experiment, we encoded the values of sensors (reward or punishment) into genes, and subjects interacted with the learning robot. Through the interaction, the subjects evaluated the robot by touching its sensors, and the robot learned appropriate combinations between input and output. Afterward, the subjects gave their scores to the experimenter, and the scores were regarded as the fitness values of the corresponding genes. These sequences were continued until the 4 generation, and then the subjects compared three of their best genes and two of the experimenter's. We found that the user-adaptive value system is suitable for the communication-robot.