Subjective Optimization

An effective way to build a gesture generator is to apply machine learning algorithms to derive a model. In building such a gesture generator, a common approach involves collecting a set of human conversation data and training the model to fit the data. However, after training the gesture generator, what we are looking for is whether the generated gestures are natural instead of whether the generated gestures actually fit the training data. Thus, there is a gap between the training objective and the actual goal of the gesture generator. In this work we propose an approach that use human judgment of naturalness to optimize gesture generators. We take an important step towards our goal by performing a numerical experiment to assess the optimality of the proposed framework, and the experimental results show that the framework can effectively improve the generated gestures based on the simulated naturalness criterion.