GIST Wins Runner-up at International Pok?mon AI Competition
AI Learns Autonomously from Limited Data
Fine-Tuning Only Essential Components from a Pre-Trained Model
The '4th Lesson' team. (From left) GIST master's students Kyungbo Kim, Eunju Kwon, Yujin Kim, and Sangyeon Park.
View original imageThe Gwangju Institute of Science and Technology (GIST) announced on December 8 that the "4th Lesson" team, consisting of graduate students from the Department of AI Convergence (master's students Kyungbo Kim, Eunju Kwon, Yujin Kim, and Sangyeon Park, supervised by Professor Kyungjoong Kim), won second place in the "Battling Track: Generation 1 Pok?mon Rules" category at the 2025 PokeAgent Challenge, an international AI competition organized by NeurIPS, the world's most prestigious artificial intelligence (AI) conference.
Both the preliminary and final rounds were held online, and the final results were announced the previous day at the NeurIPS 2025 Conference in San Diego, California, USA.
The "PokeAgent Challenge" is an international competition where AIs compete in an environment that faithfully replicates the battle mechanics of Pok?mon games (such as Pok?mon selection, move usage, and switching strategies). The competition evaluates how intelligently and strategically an AI can make decisions under uncertainty.
Participating teams train their models using self-play and data augmentation strategies within limited data and computational resources, experimenting with various tactics according to rules that reflect the Pok?mon types, moves, and stat structures that change depending on the game version.
The competition consists of two tracks: Battling and Speedrunning. The Battling track, in particular, is the core event that comprehensively evaluates an AI's ability to predict opponent strategies, plan for the long term, and respond to incomplete information.
To maximize performance despite limited resources, the GIST "4th Lesson" team adopted a fine-tuning strategy, starting with a pre-trained AI model that had already learned basic skills and additionally training only the necessary components.
To ensure more stable training, they introduced a new learning technique called the "Kron optimizer," which helps the AI learn more efficiently on its own, instead of using conventional methods.
In addition, to prevent the "loss of plasticity" problem?where the AI becomes less able to learn new information over time?they applied AID technology to continuously maintain learning capability.
The fine-tuning strategy refers to the method of performing additional training on a pre-trained model using new data or for specific tasks. This approach enables the model to adapt quickly and improve performance even in environments with limited data and computational resources, offering more efficient and stable learning compared to training a model from scratch.
The greatest strength of the "4th Lesson" team is their ability to expand data autonomously. Judging that the basic dataset of about 190,000 samples was insufficient for thorough strategy training, they built a custom experimental environment where 19 different AI models competed against each other repeatedly, generating a large volume of additional training data themselves. This enabled them to secure about 1.35 million high-quality supplementary data samples, significantly enhancing the strategic completeness of their model.
This achievement is significant not only for improving the win rate of game AIs, but also for presenting a methodology that enables reinforcement learning (where AI learns through trial and error) to work effectively even in situations with limited data or high uncertainty.
Professor Kyungjoong Kim commented, "The students achieved international success based on creative strategies and meticulous experimental design, even with limited resources," adding, "This runner-up finish demonstrates that the research and education standards of the GIST Department of AI Convergence are globally competitive."
Hot Picks Today
"Two Spend 500,000 Won in a Flash, Yet Wait Two Hours?" ... 'Dreamland' Shunned by Kids and Foreign Tourists
- The Era of Getting 200 Million Won After 30 Years Is Over... A Japanese Bank's "Perpetual Department Head" Who Just Clocked In Cries Out
- Spain Hit by Its Worst Tragedy... Lee Kangin, Moreno and Real Madrid Join Soccer World in Mourning
- They Said, “Make It Your Home for Life” - Then Suddenly Kicked 200 Passengers Off... What Happened to the Around-the-World Cruise?
- "130 Rice Bowls Sold Out in a Day"... 'Unfiltered Face' Adds 8 Million Won to Night Market Sales: China's Bungeunwoo
Student Kyungbo Kim said, "Thanks to Professor Kyungjoong Kim's passionate guidance and the support from the Department of AI Convergence, we were able to confirm the technical potential to efficiently optimize AI models even with limited resources," adding, "We will continue our research so that this technology can be applied to solving real-world industrial problems in the future."
© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.