Abstract
Representation learning (RL) has become increasingly important in machine learning which is heavily dependent on the quality of data representations. Proper RL methods can help researchers understand target objects’ nature better. From an engineering perspective, all thesis chapters revolve around innovating and using multiple RL techniques to solve real-world problems for different levels of research objects. According to whether the research objects function as a whole, the thesis divides the research objects into two main categories: individuals and systems. The thesis’ research results provide general guiding ideas for the following RL work in individual and system studies. At the individual level, the research object (e.g., a soft robot) functions as a whole. Therefore, it is necessary to study its representation by reducing or ignoring its various components’ influence. Deep neural networks (DNNs) are suitable for dealing with this situation. Soft robots have highly compliant multi-material structures with non-linear time-dependent physicochemical properties. These characteristics make controlling their flexible movements a challenging problem. This thesis first applied DNN-based models (RL method) in a batoid-inspired soft robot to achieve its accurate mechanical output control. The well-trained DNN model is an excellent robot representation form and shows the power of RL in soft underwater robot control. At the system level, the RL method starts from the system’s structure to obtain each entity’s representation rather than the whole system’s representation. Therefore, knowledge graph embedding technologies for mining structural information are more suitable for system-level RL than DNN-based models. This thesis obtains the vector representations of all the entities in the US Patent Database. It uses these representations to solve the challenge of direct computation of knowledge proximity between heterogeneous entities for the first time. Knowledge proximity is a fundamental variable in information retrieval and recommender system studies. The lack of interoperability across entities prevents the usage of current approaches from measuring knowledge proximity between heterogeneous pairs (e.g., patent-inventor) directly. This thesis initially builds ‘PatNet’ to get entity representations and shows the applications of knowledge proximity between heterogeneous entities. This pioneering RL work is significant to various patent-related applications, such as data-driven innovation, business, and policy intelligence. Inspired by the theoretical understanding of knowledge proximity measurements based on knowledge graph embedding techniques, this thesis further innovatively incorporates text embedding techniques in system-level RL research. It explores the performance of different fusion approaches in patent classification and retrieval tasks. Unlike traditional unbalanced patent retrieval methods that focus on partial aspects of patent information, this study synthesizes the textual and relational information of the patent. It demonstrates the power of RL in open-ended design opportunity exploration. All these outcomes of theories and methods applied in individual-level and system-level research objects are instructive for future research directions of data-driven RL work.