بهره‌گيري از بازنمايي گرافي در يادگيري تقويتي عميق براي بازي‌هاي متني

In recent years, Natural Language Processing (NLP) has emerged as a pivotal branch of Artificial Intelligence, particularly with the advent of large language models (LLMs) an‎d their extensive applications in text analysis an‎d knowledge inference. For future intelligent systems, the ability to comprehend an‎d reason over language, as well as to retain long-term knowledge, is essential. Consequently, integrating structured knowledge with linguistic reasoning has gained significant importance. Text-based games, due to challenges such as partial observability, vast state an‎d action spaces, sparse rewards, diverse an‎d unknown puzzles, an‎d the necessity for reasoning an‎d long-term memory, provide a compelling testbed for eva‎luating such capabilities. This thesis proposes a novel framework that employs dynamic graph representations as structured memory to enhance the guidance provided by LLMs. At each interaction step, entities an‎d relations are extracted from textual observations, encoded into graph embeddings, an‎d updat‎ed through agent-environment interactions. The LLM then generates a prior policy to constrain an‎d prioritize actions, which is incorporated into a Monte Carlo Tree Search (MCTS) using the PUCT criterion to balance exploration an‎d exploitation in a structured manner. The proposed method is implemented an‎d eva‎luated on a suite of text-based games from the Jericho platform. Quantitative results demonstrate improvements in average final rewards, sample efficiency, an‎d agent performance stability compared to prior approaches. Furthermore, qualitative analyses highlight the role of graph-based memory in preserving long-term information an‎d facilitating multi-step decision-making. However, limitations such as over-reliance on graph knowledge—potentially hindering performance in simpler, short-horizon tasks—were also identified. Overall, the findings suggest that incorporating dynamic graph representations alongside LLMs within an MCTS framework, when applied with appropriate precision an‎d safeguards, can significantly enhance the performance of intelligent agents in text-based environments, particularly in complex interactive narratives.

6

دكتر اميد روزمند از واحد آموزش عالي شهرضا مي‌باشد

149937