Ensembleinlärning för sociala robotar : Förbättrat beslutsfattande i mänskliga gruppinteraktioner

As social robots are increasingly used in various group scenarios of public services, how to make social robots be able to interact effectively in group setting is becoming a research focus for Human Robot Interaction (HRI). Compared to dyadic interaction, group social scenarios are more complex, especially the social styles of different groups are also different, which requires robots to have stronger reasoning and decision-making capabilities in groups. In recent years, some studies have introduced Graph Neural Networks (GNNs) into the field of HRI to model the relationship and interaction characteristics between interacting individuals. The Template- and Graph-based Modeling (TGM) method with Message Passing Graph Neural Network (MPGNN)as the core has made initial progress in solving the two types of decision problems: ”whom to interact with” and ”what behavior to take”. However, due to the subjectivity in human demonstration datasets, as different individuals may demonstrate varying strategies or preferences even under similar social situations, the small size of datasets, and the constraints of the models themselves, existing models still have deficiencies in generalization ability and decision accuracy. To address the above problems, this study proposed an MPGNN model optimization method based on ensemble learning. By adjusting different hyperparameter combinations and layer architectures, heterogeneous model variants were constructed and several models with their excellent performance were selected according to the Macro F1 score on the validation set. Ensemble fusion strategies such as soft voting and stacking were introduced to fuse the model outputs. The experiment was conducted on the Teenager dataset, in which teenagers controlled a robot to assist in completing group interaction tasks, and the evaluation metrics was the Macro-F1 score on the test set. The results show that the model based on ensemble learning outperforms the original baseline in the full group and the three subgroups present in the dataset, verifying the effectiveness of the proposed method. Overall, this study provides new ideas for the robust modeling of GNNs models in small sample group settings, and expands the application potential of the TGM method in real social scenarios.