Self-explaining Neural Networks in Reinforcement Learning

Sammanfattning

This thesis explores the integration of self-explaining neural networks (SENNs) into the field of reinforcement learning (RL), particularly focusing on deep reinforcement learning (DRL) models. Traditional deep learning models, while powerful, often operate as black boxes, making it difficult to interpret their decisions—a critical challenge in high-stakes domains such as network management. This work addresses the issue of model transparency by leveraging SENNs to create more interpretable DRL agents. The thesis examines the resource coordination problem in wireless mobile networks and the limitations of existing DRL methods in the context of explainability. Through a series of experiments using proximal policy optimization (PPO) and deep Q-Learning (DQN), the integration of SENNs is shown to improve interpretability without significantly sacrificing performance. The results demonstrate that incorporating interpretability into neural networks through SENNs allows agents to produce explanations of their actions, enhancing trust and usability in real-world applications. This thesis contributes to the growing field of explainable artificial intelligence (XAI) by providing a practical approach to creating self-explanatory RL agents.