Portfolio optimization, as one of the key challenges in investment management, has always been at the forefront of financial research, aiming to achieve an optimal balance between return and risk. Although Modern Portfolio Theory (MPT) laid the foundation for optimization by introducing the mean-variance framework, the complexities of modern markets, including dynamic volatility, extreme risks, and nonlinear dependencies, have highlighted the need for more advanced methods. This study leverages a dynamic copula model to assess time-dependent asset dependencies and integrates multi-objective criteria, including Worst-Case Conditional Value at Risk (WCVaR) and risk-adjusted return, to propose two optimization frameworks based on the Particle Swarm Optimization (PSO) algorithm and Generative Adversarial Networks (GAN). The methodological core of this research is a systematic comparison of the capabilities of PSO and GAN in identifying optimal portfolios. While PSO explores the solution space using a swarm intelligence mechanism, focusing on simultaneously improving the Sharpe ratio and reducing WCVaR, GAN employs generative and discriminative networks to simulate complex market patterns and design portfolios with greater resilience to crisis conditions. The empirical data for this study is based on historical information from companies listed on the Tehran Stock Exchange, processed and analyzed in the Python environment. Key findings indicate that both PSO and GAN models significantly outperform classical methods such as Markowitz and equal-weighted portfolios. However, GAN demonstrates superior performance by improving the Sharpe ratio while maintaining WCVaR levels comparable to those of the PSO model. This performance gap stems from GAN's ability to model nonlinear relationships and identify assets with negative correlations in volatile market conditions.