2025, Vol.30, No.9

Journal of The Korea Society of Computer and Information 2024 KCI Impact Factor : 0.81

pISSN : 1598-849X / eISSN : 2383-9945
https://journal.kci.go.kr/jksci

Deep Q-Learning (DQL) is a representative deep reinforcement learning method. Deep Q-Network (DQN), based on DQL, has improved the performance by replaying past experiences stored in a replay memory. In the DQN, past experiences are randomly sampled from a single replay memory for training. This paper proposes a heuristic-based approach using multiple replay memories to further enhance the performance of DQL. Specifically, experiences are categorized based on knowledge of the target problem and stored in their corresponding replay memories. During training, a fixed number of experiences are randomly sampled from each replay memory. The proposed method was applied to the Cart Pole, the Mountain Car, and the Ball Avoider problems, and comparative experiments were conducted to compare with existing methods. The experimental results show that the proposed method can significantly improve the performance of DQL.

A number of studies have been conducted using diverse data sources to forecast cryptocurrency prices. However, unlike the rational behavior of human investors who manage risk through asset diversification between risky and safe assets, most deep learning models for cryptocurrency price prediction have overlooked features related to safe assets that may not appear directly connected to Bitcoin as risky assets. In response, this study demonstrates that a deep learning price-forecasting model designed by incorporating safe assets—specifically Korean bond indices—alongside Bitcoin significantly improves prediction performance. In our experiment, features derived from the Korean Bond Indexes were selected through the Granger Causality Test to assess their relevance for Bitcoin price prediction. To ensure reliability in forecasting, the dataset was preprocessed to achieve stationarity across all features. The results indicate that a model trained on both Bitcoin-related variables (representing risky assets) and Korean Bond features (representing safe assets) delivers enhanced predictive performance.

This study designed and implemented an ethical Q&A chatbot using the Retrieval-Augmented Generation (RAG) technique to support the ethical use of generative AI. Based on the evaluation using the PDF document "Generative AI Ethics Guidebook (2023)", the RAG-based chatbot demonstrated superior response performance in handling unstructured data (such as tables and diagrams) within the document compared to existing generative AI models (ChatGPT, Gemini, DeepSeek, etc.). Using the RAGAS(Retrieval-Augmented Generation Assessment Score) evaluation framework, key metrics— including Context Recall and Factual Correctness—were analyzed. The results confirmed that the RAG chatbot enhances the reliability of document-based information retrieval and can be effectively used in AI ethics education. This research provides a foundational methodology for developing AI ethics guidelines and applying them in educational settings.