Visual emotion recognition, one of the most critical skills connecting the relationship among human society, aroused extensive attention in human-robot interaction (e.g., Matteo, 2020). Nowadays, some deep neural network (DNN) based models trained with human facial expression images can recognize basic human facial emotions with high accuracy (see Li & Deng, 2018). For both engineering and vision science, it is interesting to clarify the differences between machine recognition models and humans in recognizing facial expressions made by artificial agents. Our study used Nikola, a FACS-based robot with 35 degrees of freedom on its face, to make a 3D stimulus generating human-like facial expressions. We chose a set of Nikola's action-unit (AU) parameters for each of the seven basic facial expressions (Anger, Disgust, Fear, Happiness, Sadness, Surprise, Neutral) in two ways. (A) We adjusted the AU parameters for prototype expressions based on Ekman's theory (Sato et al., 2022). (B) Using a Bayesian Optimization algorithm, we found AU parameters that the corresponding expression image received the maximum rating by Py-Feat, a DNN model for human expression classification. We then asked forty human participants, aged between 18 and 40 years, to evaluate the robot expressions made by the two methods using a 7-point rating task. The results showed that Py-Feat-based optimization outperformed prototype expressions for Anger, Disgust, Sadness, and Surprise (p<0.001), which suggests the effectiveness of the DNN-based model in expression recognition on Nikola's facial expressions. However, Py-Feat-based optimization could not achieve higher ratings for Happiness and Fear expressions, despite the Py-Feat rating predicting significant improvements. The observed difference between Py-Feat scores and human scores in evaluating certain facial expressions made by a human-like agent reveals a difference in processing strategy in facial expression recognition between humans and a popular DNN-based model.