[AML]Pokémon Battle Agent based on LLMs
Keywords: LLMs, Agent, SFT, Knowledge-Augmented Generation, Self-Consistency
TL;DR: We plan to develop a Pokémon battle agent based on LLMs.
Abstract: The rapid development of LLMs has led to widespread applications in interactive environments, particularly in gaming, where LLM agents demonstrate impressive decision-making and strategy execution capabilities. We focus on developing a Pokémon battle agent based on GLMs through combining four techniques to improve agent's contextual understanding and generate effective battle commands: (1) In-context reinforcement learning; (2) Knowledge-augmented generation; (3) Consistent action generation; (4) Supervised Fine-Tuning. Through experiments on Pokémon Showdown with robots and human, we evaluate the model's win rates and strategic performance, aiming to contribute to the development of LLM agents capable of dynamic and complex environments. Results showed that despite the capability limitations of LLMs, they still demonstrated acceptable performance, proving their potential ability in analyze and strategize in the complex domain of Pokémon battles.
Submission Number: 23
Loading