YJYan Jin

Yan Jin

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Yan Jin has a total of 48 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Knowledge capture, adaptation and composition (KCAC): A framework for cross-task curriculum learning in robotic manipulation

    by Xinrui Wang, Yan Jin

    Reinforcement learning (RL) has demonstrated remarkable potential in robotic manipulation but faces challenges in sample inefficiency and lack of interpretability, limiting its applicability in real world scenarios. Enabling the agent to gain a deeper understanding and adapt more efficiently to diverse working scenarios is crucial, and strategic knowledge utilization is a key factor in this process. This paper proposes a Knowledge Capture, Adaptation, and Composition (KCAC) framework to systematically integrate knowledge transfer into RL through cross-task curriculum learning. KCAC is evaluated using a two block stacking task in the CausalWorld benchmark, a complex robotic manipulation environment. To our knowledge, existing RL approaches fail to solve this task effectively, reflecting deficiencies in knowledge capture. In this work, we redesign the benchmark reward function by removing rigid constraints and strict ordering, allowing the agent to maximize total rewards concurrently and enabling flexible task completion. Furthermore, we define two self-designed sub-tasks and implement a structured cross-task curriculum to facilitate efficient learning. As a result, our KCAC approach achieves a 40 percent reduction in training time while improving task success rates by 10 percent compared to traditional RL methods. Through extensive evaluation, we identify key curriculum design parameters subtask selection, transition timing, and learning rate that optimize learning efficiency and provide conceptual guidance for curriculum based RL frameworks. This work offers valuable insights into curriculum design in RL and robotic learning.

  • A Method for Synthesizing Ontology-Based Textual Design Datasets: Evaluating the Potential of Large Language Model in Domain-Specific Dataset Generation

    by Yunjian Qiu, Yan Jin

    In engineering disciplines, leveraging generative language models requires using specialized datasets for training or fine-tuning the preexisting models. Compiling these domain-specific datasets is a complex endeavor, demanding significant human effort and resources. To address the problem of domain-specific dataset scarcity, this study investigates the potential of generative large language models (LLMs) in creating synthetic domain-specific textual datasets for engineering design domains. By harnessing the advanced capabilities of LLMs, such as GPT-4, a systematic methodology was developed to create high-fidelity datasets using designed prompts, evaluated against a manually labeled benchmark dataset through various computational measurements without human intervention. Findings suggest that well-designed prompts can significantly enhance the quality of domain-specific synthetic datasets with reduced manual effort. The research highlights the importance of prompt design in eliciting precise, domain-relevant information and discusses the balance between dataset robustness and richness. It is demonstrated that a language model trained on synthetic datasets can achieve a level of performance comparable to that of human-labeled, domain-specific datasets in terms of quality, offering a strategic solution to the limitations imposed by dataset shortages in engineering domains. The implications for design thinking processes are particularly noteworthy, with the potential to assist designers through GPT-4's structured reasoning capabilities. This work presents a complete guide for domain-specific dataset generation, automated evaluation metrics, and insights into the interplay between data robustness and comprehensiveness.


  • Impact of Task Constraint on Agent Team Size of Self-Organizing Systems Measured by Effective Entropy

    by Hao Ji, Yan Jin

    Self-organizing systems can perform complex tasks in unpredictable situations with adaptability. Previous work has introduced a multiagent reinforcement learning-based model as a design approach to solving the rule generation problem with complex tasks. A deep multiagent reinforcement learning algorithm was devised to train self-organizing agents for knowledge acquisition of the task field and social rules. The results showed that there is an optimal number of agents that achieve good learning stability and system performance. However, finding such a number is nontrivial due to the dynamic task constraints and unavailability of agent knowledge before training. Although extensive training can eventually reveal the optimal number, it requires training simulations of all agent numbers under consideration, which can be computationally expensive and time consuming. Thus, there remains the issue of how to predict such an optimal team size for self-organizing systems with minimal training experiments. In this article, we proposed a measurement of the complexity of the self-organizing system called effective entropy, which considers the task constraints. A systematic approach, including several key concepts and steps, is proposed to calculate the effective entropy for given task environments, which is then illustrated and tested in a box-pushing case study. The results show that our proposed method and complexity measurement can accurately predict the optimal number of agents in self-organizing systems, and training simulations can be reduced by a factor of 10.


  • Open Access

    Exploring CausalWorld: Enhancing robotic manipulation via knowledge transfer and curriculum learning

    by Xinrui Wang, Yan Jin

    This study explores a learning-based tri-finger robotic arm manipulating task, which requires complex movements and coordination among the fingers. By employing reinforcement learning, we train an agent to acquire the necessary skills for proficient manipulation. To enhance the efficiency and effectiveness of the learning process, two knowledge transfer strategies, fine-tuning and curriculum learning, were utilized within the soft actor-critic architecture. Fine-tuning allows the agent to leverage pre-trained knowledge and adapt it to new tasks. Several variations like model transfer, policy transfer, and across-task transfer were implemented and evaluated. To eliminate the need for pretraining, curriculum learning decomposes the advanced task into simpler, progressive stages, mirroring how humans learn. The number of learning stages, the context of the sub-tasks, and the transition timing were found to be the critical design parameters. The key factors of two learning strategies and corresponding effects were explored in context-aware and context-unaware scenarios, enabling us to identify the scenarios where the methods demonstrate optimal performance, derive conclusive insights, and contribute to a broader range of learning-based engineering applications.

  • Document Understanding-Based Design Support: Application of Language Model for Design Knowledge Extraction

    by Yunjian Qiu, Yan Jin

    Design knowledge in the vast amount of design reports and documents can be an excellent resource for designers in their practice. However, capturing such domain-specific information embedded in long-length unstructured texts is always time-consuming and sometimes tricky. Therefore, it is highly desirable for a computer system to automatically extract the main knowledge points and their corresponding inner structures from given documents. In this study of document understanding for design support (DocUDS), a design-perspective knowledge extraction approach is proposed that uses phrase-level domain-specific labeled datasets to finetune a Bidirectional Encoder Representation from Transformers (BERT) model so that it can extract design knowledge from documents. The BERT model finetuning attempts to blend in the domain-specific knowledge of well-recognized domain concepts and is based on the datasets generated from design reports. The model is utilized to map the captured sentences to the main design entities , , and . In addition, this approach uncovers inner relationships among the sentences and constructs overall structures of documents to enhance understanding. The definitions of design perspectives, inter-perspective relations, and intra-perspective relations are introduced, which together capture the main design knowledge points and their relations and constitute an understanding of the design domain knowledge of a text. The case study results have demonstrated the proposed approach's effectiveness in understanding and extracting relevant design knowledge points.


  • Long-Range Risk-Aware Path Planning for Autonomous Ships in Complex and Dynamic Environments

    by Chuanhui Hu, Yan Jin

    Path planning and collision avoidance are common problems for researchers in vehicle and robotics engineering design. In the case of autonomous ships, the navigation is guided by the regulations for preventing collisions at sea (COLREGs). However, COLREGs do not provide specific guidance for collision avoidance, especially for multi-ship encounters, which is a challenging task even for humans. In short-range path planning and collision avoidance problems, the motion of target ships is often considered as moving at a constant velocity and direction, which cannot be assumed in long-range planning and complex environments. The research challenge here is how to factor in the uncertainty of the motion of the target ships when making long-range path plans. In this paper, we introduce a long-range path planning algorithm for autonomous ships navigating in complex and dynamic environments to reduce the risk of encountering other ships during future motion. Based on the information on the position, speed over ground, and course over ground of other ships, our algorithm can estimate their intentions and future motions based on the probabilistic roadmap algorithm and use a risk-aware A* algorithm to find the optimal path that has low accumulated risk of encountering other ships. A case study is carried out on real automatic identification systems (AIS) datasets. The result shows that our algorithm can help reduce multi-ship encounters in long-term path planning.


  • Exploring Visual Cues for Design Analogy: A Deep Learning Approach

    by Zijian Zhang, Yan Jin

    The goal of this research is to develop a computer-aided visual analogy support (CAVAS) framework to augment designers’ visual analogical thinking by stimulating them by providing relevant visual cues from a variety of categories. Two steps are taken to reach this goal: developing a flexible computational framework to explore various visual cues, i.e., shapes or sketches, based on the relevant datasets and conducting human-based behavioral studies to validate such visual cue exploration tools. This article presents the results and insights obtained from the first step by addressing two research questions: How can the computational framework CAVAS be developed to provide designers in sketching with certain visual cues for stimulating their visual thinking process? How can a computation tool learn a latent space, which can capture the shape patterns of sketches? A visual cue exploration framework and a deep clustering model CAVAS-DL are proposed to learn a latent space of sketches that reveal shape patterns for multiple sketch categories and simultaneously cluster the sketches to preserve and provide category information as part of visual cues. The distance- and overlap-based similarities are introduced and analyzed to identify long- and short-distance analogies. Performance evaluations of our proposed methods are carried out with different configurations, and the visual presentations of the potential analogical cues are explored. The results have demonstrated the applicability of the CAVAS-DL model as the basis for the human-based validation studies in the next step.


  • Path Planning for Autonomous Systems Design: A Focus Genetic Algorithm for Complex Environments

    by Chuanhui Hu, Yan Jin

    Path planning has been a hot research topic in robotics and is a vital functionality for autonomous systems. As the time complexity of traditional path planning algorithms grows rapidly with the complexity of the problem, evolutionary algorithms are widely applied for their near-optimal solutions. However, evolutionary algorithms can be trapped in a local optimum or converge to infeasible solutions, especially for large search spaces. As the problem scale increases, evolutionary algorithms often cannot find feasible solutions with random exploration, making it extremely challenging to solve long-range path-planning problems in environments with obstacles of various shapes and sizes. For long-range path planning of an autonomous ship, the current downsampling map approach may result in the disappearance of rivers and make the problem unsolvable. This paper introduces a novel area-based collision assessment method for genetic algorithm (GA) that can always converge to feasible solutions with various waypoints in large-scale and obstacle-filled environments. Waypoint-based crossover and mutation operators are developed to allow GA to modify the length of the solution during planning. To avoid the premature problem of GA, the mutation process is replaced by a self-improving process to let the algorithm focus the operations on any potential solutions before discarding them in the selection process. The case studies show that the proposed GA-focus algorithm converges faster than RRT* and can be applied to various large-scale and challenging problems filled with obstacles of different shapes and sizes, and find high-quality solutions.