World Models Unlock Optimal Foraging Strategies in Reinforcement Learning Agents
This study investigates the computational mechanisms underlying the “when-to-leave” decision in biological patch foraging—a core ecological decision—and leverages these insights to design more interpretable, biologically grounded AI decision models. We propose a model-driven reinforcement learning framework integrating sparse world modeling, predictive representation learning, and forward simulation. We provide the first theoretical proof and empirical validation that agents equipped with a learned world model autonomously converge to the biologically optimal patch-leaving strategy predicted by the Marginal Value Theorem (MVT), without explicit programming. Critically, their decisions are driven by minimization of prediction error—not merely reward maximization. Compared to conventional model-free RL, our approach significantly improves fidelity to observed foraging behavior, ecological plausibility, and decision interpretability. This work establishes a novel paradigm bridging computational neuroscience and trustworthy AI.