MENOMONEE, Wis. – The University of Wisconsin-Stout’s 360-degree AI education approach prepares graduates to meet the needs of a rapidly evolving workforce by incorporating artificial intelligence training into all degree programs. The final projects of two groups of Applied Mathematics and Computer Science students in the Advanced Machine Learning course resulted in a model that positively impacts educational institutions and one that creates highly competitive AI opponents for the classic board game.
One group built the RAG model (Search Augmented Generation) to help students search UW-Stout course bulletins in a more engaging way. Another company created a digital version of the board game Othello by training a model using Monte Carlo Tree Search (MCTS) and a database of top international players.
“The motivation for this class is for students to take cutting-edge papers, read them, understand them, and implement them. Five years from now, when they’re in their careers, they’ll know how to read a technology-heavy mathematics paper and translate it into code,” said AMCS Program Director. Professor Seth Dutter. “These are the types of top-level projects that I want to give my students. The experience defines them.”
Especially. It’s a polytechnic, and it goes to the heart of UW-Stout. ”
of Stout News RAG
When developing a group project, tyler smithof Rochester. Matthew Peplinskiof Milwaukee. aaron kingof Rhinelander. and Kyler Nikolaifrom Rochester wanted to create an easy way to gather information about UW-Stout courses, degree programs, minors, and certifications without having to read the entire course bulletin.
“We created the Stout News RAG, which allows us to take questions about Stout courses and programs and get back accurate information within seconds,” Smith said. “New students can use this program when applying to Stout to see what classes and programs are offered, and current students can use it to schedule classes for next semester. This program actively uses ChatGPT to curate answers that provide the information you need in the right way. And because the data comes directly from current Stout bulletins, ChatGPT It provides more specific information, so there’s no chance of finding out-of-date or irrelevant information.”
The RAG model allows users to describe what they are looking for in a sentence format, rather than just using a single word search. Smith explained that the model consists of three parts: a prompt, data acquisition, and a response from the LLM (Large-Scale Language Model). A prompt is something a user, such as a student, asks the model. The model then searches the stored data to find which sentences, paragraphs, or text documents are most similar to the question being asked.
“This is determined by the embedding model. An embedding model is, simply put, a specific model that is trained to find similarities between texts using a vector space,” he said. “The RAG then takes the top results for the most similar text and throws it into a new prompt. This prompt includes all the details of the question asked, the information retrieved, and any rules in place, so when the model retrieves an LLM, such as ChatGPT, to summarize the information, the LLM won’t make up information it doesn’t have.”
Peplinski added that the point of RAG is to leverage untrained information to leverage the inference capabilities of existing LLM models. “We used two different AI models in this process: one for text embedding and one for summarization,” he said. “Embedded models were used to determine how similar a user’s question was to the information available in our dataset. We then extracted the top five bits of most similar information and sent it to the LLM (ChatGPT in our case). The LLM was only used to provide an overview of the information we provided and was strictly instructed not to fabricate false information.”
Peplinski and Smith thought the best part of this project was seeing how thorough the RAG model was during the final class presentation, which answered the prompts and assignments from the participants almost perfectly.
The most difficult part was collecting all of UW-Stout’s program information and automating the data collection process so that it didn’t have to be exported manually.
Smith plans to graduate this spring and begin working at Federated Insurance this summer. He feels this project helped prepare him for the industry, as many companies already use in-house chatbots to search for documents. “They are very likely using a RAG model structure similar to the one we created.
It helps to come into the industry with that knowledge,” he said.
Dutter agreed, adding, “Companies will benefit from employees who know how to code RAGs. This is a useful skill to help companies that want to be able to search for information in documents, sometimes spanning decades.”
Othello A.I. project
After 4 days, nathan lacrosseof Portage. Noah Stitgenof Roddy. jake swansonfrom Eden Prairie, Minnesota. and Lindsey Redepeningof Elk River, Minn., programmed a digital version of Othello from scratch, creating the interface and teaching the AI model the parameters of play, how to take the opponent’s pieces, and what it takes to win the game. Over the next two weeks of the project, they optimized the game to create the most challenging and ultra-aggressive AI opponents possible.
download Games on GitHub.
Lacrosse originally wanted to choose chess for the group’s project, but chose Othello, a board game that was simpler and easier to watch and play. Also, unlike checkers, Othello cannot end in a draw.
The group used MCTS, an ultra-fast algorithm used in AI decision-making processes, specifically games, to develop the model. We simulated 3,500 Othello games per second. “When evaluating board position, MCTS runs a number of random games and selects the move that statistically leads to a win,” LaCrosse said.
They tried exploring the tree to make sure it worked properly. “In fact, we found that the game was programmed incorrectly because the algorithm was picking up on false moves that gave it an unfair advantage,” LaCrosse said.
We then added a neural network to the tree search that studied a database of top French Othello players. They used the database to train the network.
The group had fun competing against every variation of gameplay the model allowed. For the published version, LaCrosse said tree search was chosen because it had the most unique playstyle.
“This project familiarized us with the process of creating a complete application and prepared us for industry. We all worked on different elements of the program and were able to put it all together in the end,” said LaCrosse, who is pursuing a Ph.D. in the field of computer science.
UW-Stout, members of universities in wisconsinteeth Wisconsin Institute of Technologywith an emphasis on applied learning, business and industry collaboration, and career outcomes. For more information, focus 2030 Strategic planning.
