My Research is about how to leverage large language model (LLM) to empower intelligent agents for their capacity of extracting plan information related with task goal, then executing plans and further refining them in an optimal way, which is a subarea of Artificial General Intelligence (AGIs). Currently, my primary work is to build intelligent robots on Android platform, which can take relevant task information from different sources, enhance information based on execution requirement and drive them to take actions accordingly. I have built up an open-source data sampling platform Android Online for Plan to allow annotators to finish task completion following refined instructions from search engine. Inspired by other highly-relevant researches in Android Agent area, I'm now building an end-to-end solution in order to automate plan retrieval, refinement and drive Android robots to make end-to-end executions.
CONOP Optimization Biochronology as a Traveling Salesman Problem, which can be solved by a simulated annealing process. However, inside the inner loop, there exists a non-convex restriction that leads to the major calculation bottleneck. In this research, I focuses on developing a visual studio-based plugin to help us trace actual performance curve aligning with the theoretical analysis. After such alignment, dynamic programming calculors together with CPU caches technology has been applied for optimization of the sequential model. Meanwhile, a consistent sampling MDP model has been given to speed up the program, while a heuristic ML model has been used to dynamically swith on/off multi-CPU parallelization. Furthermore, for the purpose of accelerating the algorithm, a non-consistent MDP model has been proposed to better random walk in the solution space and primiarly verfied.
Establish a Reinforcement Learning (RL) Model to generate recommended product items, so our system can reach the balance between gaining product profit and satisfying customer interests. The main problem is how to setup RL model four tuples (state, action, reward, transition) and integrate RL model with system.
Conference paper for ICRA 2015. Protein-ligand prediction plays a key role in drug discovery. Nevertheless, many algorithms are over reliant on 3D structure representations of proteins and ligands which are often rare. Techniques that can leverage the sequence-level representations of proteins and ligands are thus required to predict binding affinity and facilitate the drug discovery process. We have proposed a deep learning model with an attention mechanism to predict protein-ligand binding affinity. Our model is able to make comparable achievements with state- of-the-art deep learning models used for protein-ligand binding affinity prediction.
In this paper, the process of implementing the JXME’s MIDP protocol and HTTP connection between relay proxy and mobile peer are introduced. In the meanwhile, the encapsulation of Byte Streams on the end of mobile is analyzed. Fixing with several serious programming bugs that severely impact the stability and scalability of MIDP framework, one basic programming framework for sharing resources among mobile peers is proposed.
This research is trying to estalbish family Graph given the inheritance disease records given by Huaxi Hosptal. One RDF-based framework has been introduced to model entity-entity relations and link relevant resources accordingly. We explore the graph via one DFS algorithm with several adaptions to avoid deadlock loop during exploration and analyze the probability of the gene reproducing path in the graph.
Research advisor: Professor Jiliu, Zhou
This summary has been given out for all publications about my previous research work sorted chronically.