Minesweeper Implementation using Deep Q-Network (DQN)
[Code]
- Implemented most of the game environment entirely from scratch using functions.
- Significantly improved win rate and cumulative score through effective normalization.
- Designed a concrete and fine-grained reward system tailored to the Minesweeper environment.
- Achieved a training win rate of 65% and a test rate of 56%.
School Announcements Classification Task
[Code]
- Designed a text classification task to categorize announcements relevant to students’ careers, such as scholarships, extracurricular activities, and employment.
- Crawled approximately 10,000 text entries from the web and manually labeled them to construct a custom dataset.
- Fine-tuned Multilingual BERT and KLUE-BERT models, achieving over 0.9 both accuracy and F1 score.
Bank Deposit Subscription Classification Task
[Code]
- Worked on a binary classification task with a highly imbalanced dataset.
- Applied various machine learning techniques such as oversampling (SMOTE), Bayesian optimization, and K-Nearest Neighbors, using frameworks like LGBMClassifier.
- Explored multiple approaches to optimize performance, including standardization, PCA, and stepwise method.
