A nation, although for dynamic and adaptable response to the diverse needs of industry, park planning and operation Tech, I want to emphasize that we hope to answer the following questions of users in real time, accurately and scientifically, like if I want to open like kindergarten and then in the zhangjiang high tech Park in Shanghai. Can you recommend a suitable address or question like, I want to start an artificial intelligence company? Could you recommend some suitable industry Park in Shanghai? If that were questions like this from our users, then I'm going to hand over to Dr Suchi to present the detail of This project. Suchi, the floor is yours.
You mute. You
I don't hear anything yet. Oh, I see you're checking here. Go ahead. Oh,
okay, so my name is Wang, and I'm a PhD student at Hong SIU University and a Hong Kong Polytechnic University. And I'm very glad to be here to share our research outcomes in which we attempt to integrate the large language model into the planning and operation of the industrial parks that we call the Ippo. And we know that the emergence of the LM presents opportunity, but also the issues. And for instance, if a user query about opening a kindergarten in jeonjuang, as Professor Liu mentioned before, and due to the missing specific data, might only generated the general considerations, it could even produce a non existent address. Moreover, for the brand location recommendation, like deciding to open the 711 may not check the vicinity for the companion stores because it lacks the spatial Association analysis capability. So let's see how we address the challenges integrating the for the Ippo solutions. First is how to construct a comprehensive dataset. So we introduce the industrial school keygen. It's a multi level, large, large scale knowledge graph for the industrial parks in Shanghai. And second, adapting the for the industrial park Knowledge Graph work was such a hurdle. So the industrial park Knowledge Graph have diverse data types, needing the near real time geospatial data handling. So our solution, industrial scoop, GBD, enables LMS to dynamically adapt to the knowledge graph database with RAC and other geospatial tools. And the third refers to the flexibility and interpretability and we enhance the LM multi step rezoning capability through the monotrez search and agent to improve this. And we define the industrial school kg as graph. G equals E, S, y and E represents the energy that encompasses elements such as companies grades and industrial parks, and the relational triples denote the relationships between the entities, including the company located in industrial Park, and other connections like the adjacency and similarity and the attribution of triples provide the details on the attributes of each entity, and the table shows the details of the statistic information. And to build the industries group, TG, we follow a five step construction pipeline, and we start with a spatial temporary data collection and feature extraction. Now the features are done especially join mapping to the scales of the 100 meter grades and industrial parks. And following this, we constructed the relation relational triples and attributional triples covering the three dimensions. And finally, we use a new four Gen knowledge graph database for the management and for the methods we developed the industry. GPT IS LM travel agent and combines the tool interaction with a graph database following the React style and to optimize the decision making, we use a monocular tree search, and I will show how It works with the following example. So our framework in the diagram toggles the typical signing query, and then consider that I want to open the Bank of China near Chenxi road. Please recommend specific address, and then at the root node, the industrial scope, GBT decomposes the query into the first sub question, it might consider searching for the nearby grades, and then concurrently invokes the potential towards like a server searcher or geo encoder to interact with the database and battery map to Geo located The mesh and address to the specific grades, then our LM will reflect and score the tour execution result to determine if a solution has been found, and the new child nodes are then generated from the feedback it might invoke the rank faster to sort The identified grades based on their features leading to the top five grades, and in the next steps, the Monte Carlo Tree picks the best known for further expansion, and the top five grade are then decoded, resulting the five rail address for The recommendation for the user. And these are tools we designed to support a specific task, and we combine this to generate the rezoning chart that adapt to the barrier test. And for our experiment, we test the GPT four with various data inputs such as standalone and the table and the table and result shows that the GPT four perform will with the industrial school PG. We also compare the industrial school GPT with other prompting methods such as CLT, and react to pro H performance and to assess how, how can industry scoop DVDs of the industrial part functional planning? We also compare with the model like the like JBM and GCN, and use the hill numbers for the diversity measurement and the result on the three cases highlights the industrial scope, gbts capability, okay, thank you for your attention. Well, appreciate if you have any discussions about our work. Thank you
wonderful. Thank you. Leon, thank you. CG, both of their emails were shared in the city. Science network email. So if you want to continue the conversation, please reach out to them and purpose of time, we will pass right on to project. Two project. Two is the pro social urban development.
