[4L1-GS-5g] Agents

9:15 AM - 9:30 AM JST(12:15 AM - 12:30 AM UTC)

[4L1-GS-5g-02] Automatic Design of Predator–Prey Tactics via LLM-Based FSM Generation and Coevolutionary OptimizationA Unified Framework Integrating Tactical Log Analysis and Meta-Evolutionary Improvement Loops

〇Masaaki Isozaki¹ (1. *)

Keywords:

Tactical FSM Generation with LLM,Coevolutionary optimization of predator–prey systems,Multi-Agent Simulation,Artificial Life

In predator–swarm simulations, tactical design based on finite state machines (FSMs) often suffers from redundancy and the difficulty of manual tuning. This study proposes a meta-evolutionary loop that integrates tactical FSM generation using large language models (LLMs), behavioral parameter optimization via genetic algorithms (GAs), and LLM-based strategy refinement driven by tactical log analysis. Using predefined tactical categories for predators and prey, we compare the initial FSMs, GA-optimized FSMs, and LLM-refined FSMs, evaluating tactical events and changes in key performance metrics. Analyses of FSM structural metrics, behavioral parameters (genomes), and performance indicators show that the proposed framework simultaneously improves FSM structural organization and behavioral effectiveness, enabling autonomous refinement of tactical behaviors.