Language Agents for Interactive Fiction Text Games: Leveraging Large Language Models to Navigate Hard Exploration Problems

Abstract

Interactive fiction text games are a valuable sandbox for research on AI agents, since they require elaborate planning, reasoning, and strategizing. Complex games have significantly large action spaces, high variability, and sparse rewards, which all pose a challenge to Reinforcement Learning (RL) methods previously proposed for game playing. Therefore, the goal of this thesis is to come up with an interpretable and generalizeable language-agent-based approach for playing interactive fiction text games. The thesis explores tree-searching and Large Language Model (LLM)-generated reasoning and learning, allowing the agent to explore the states of its environment and learn from its experience while also leveraging the semantic reasoning abilities of LLMs to efficiently navigate the search space. We use GPT-4 Turbo as the primary model powering the agent and use the games in the Jericho interface developed by Microsoft [2] as the benchmark for evaluating performance. Although the results we obtained currently do not surpass those of the state-of-the-art agent for text games [11], our proposed modules are more generalizeable and require no training time; at the same time, they achieve relatively higher scores compared to previous language agents playing Jericho games, suggesting the potential promise of these methods for complex exploration problems.