Explainable Deep Reinforcement Learning for Knowledge Graph Reasoning
Di Wang · Advances in computational intelligence and robotics book series · 2023
Artificial intelligence faces a considerable challenge in automated reasoning, particularly in inferring missing data from existing observations. Knowledge graph (KG) reasoning can significantly enhance the performance of context-aware AI systems such as GPT. Deep reinforcement learning (DRL), an influential framework for sequential decision-making, exhibits strength in managing uncertain and dynamic environments. Definitions of state space, action space, and reward function in DRL directly dictate the performances. This chapter provides an overview of the pipeline and advantages of leveraging DRL for knowledge graph reasoning. It delves deep into the challenges of KG reasoning and features of existing studies. This chapter offers a comparative study of widely used state spaces, action spaces, reward functions, and neural networks. Furthermore, it evaluates the pros and cons of DRL-based methodologies and compares the performances of nine benchmark models across six unique datasets and four evaluation metrics.