The Japanese Society for Artificial Intelligence

This paper focuses on PokeLLMon, a large language model (LLM)-based agent for efficient Pokemon battles. To mitigate the issues of PokeLLMon in long-term battles, we propose three prompt engineering techniques aimed at enhancing strategic decision-making. The first technique ensures that the agent accounts for incoming damage, enabling adaptive responses. The second and third involve stepwise inference of the battle situation and strategy, which improve the agent's inference ability by estimating the opponent's potential behavior and the current situation. Experimental results using GPT-4 have shown that these techniques significantly enhance performance, as evidenced by improved win rates and higher scores in long-term battles.