This work introduces a framework that leverages multimodal learning and reinforcement learning to evaluate in‑game sound designs efficiently. Sound design heavily influences gameplay and accessibility, yet existing objective evaluation typically requires retraining an agent from scratch for each sound variant. We first pre‑train a multimodal encoder that projects visual and auditory streams into a shared latent space. A reinforcement‑learning agent is then trained on this latent representation. For each new sound design we fine‑tune only the encoder while keeping the policy network fixed, thus isolating the impact of sound changes on agent behavior. Experiments across several game genres compare blind agents (using sound input alone) under different sound designs. Results show the method can detect performance shifts caused by degraded or altered audio and reduces computational cost dramatically compared with full retraining, offering a practical and scalable solution for sound‑design evaluation. The source code, data, and supplementary documents are publicly available at https://github.com/thainv0212/tog-reusing-agents.
Review