Bridging GANs and Reinforcement Learning: Proximal Policy Optimization for Enhanced Design Outputs

Generative Adversarial Networks (GANs) have demonstrated significant potential in floor plan generation by producing layouts with diverse and structurally coherent understanding. However, the House-Gan++ model often struggles with stability and constraint adherence due to the adversarial nature of their training process. This research introduces a novel hybrid training framework integrating Proximal Policy Optimization (PPO) and GANs to enhance generative performance and stability. The proposed method integrates PPO into the GAN training loop to enhance generator loss optimization through non-adversarial feedback, producing realistic and structurally valid architectural plans. The PPO agent navigates the generator through a policy network that modifies latent abstract features using advantage estimates. The approach enhances the adversarial learning stability by stopping mode collapse and improves how the generative model follows user-defined spatial constraints. Training occurs through the RPLAN dataset because it contains a substantial collection of professional single-story house plans, which ensures great architectural variety and authenticity. Experimental tests between PPO-GAN and House-GAN++ evaluate their performance through four training loss metrics: Generator Loss, Discriminator Loss, PPO Loss, and L1 Loss and two objective quality metrics: Fréchet Inception Distance (FID) and Graph Edit Distance (GED). Including PPO produces layouts that achieve steady convergence, superior spatial constraint compliance, and reduced FID scores, proving their improved authentic quality. The PPO-GAN model demonstrates higher computational expenses and training instability at first but proves reinforcement learning (RL) as an effective GAN-based architectural design enhancement through its superior long-term performance and stability. This study aims to contribute to AI-driven generative design by bridging GANs and RL.