Objective-Aware Representation Learning for Language-Instructed Procedural Level Generation via Reinforcement Learning

Abstract

Procedural content generation via reinforcement learning enables controllable game content generation by training an agent to optimize desired content properties through reward functions. However, existing controllable approaches mainly rely on predefined numerical conditions, which makes it difficult to reflect high-level and compositional design intent expressed in natural language. This thesis investigates natural language instructions as an intuitive control interface for reinforcement learning-based procedural content generation. This thesis proposes a language-conditioned generation framework that transforms natural language instructions into latent representations and uses them to guide the decision-making process of a reinforcement learning agent. The framework first supports text-based control for single-objective instructions and is then extended to multiple-objective instructions through structured representation learning. By separating instruction semantics into objective-specific factors and learning both objective activation and target intensity, the proposed approach reduces interference among multiple objectives and enables more reliable control under complex instructions. Experimental results show that objective-specific language representations improve controllability compared with general-purpose sentence representations. In addition, the structured multiple-objective representation improves control under complex instructions while maintaining single-objective performance and generalizing to reduced objective combinations not directly used as training compositions. These results demonstrate that natural language can serve as an expressive and practical control modality for procedural content generation.