Dynamic Multi-Agent Reinforcement Learning for Competitive E-Commerce Price Optimization: Curriculum-Driven SAC, Equilibrium Analysis and Ethical Safeguards

We present a comprehensive study on dynamic multi-agent reinforcement learning (MARL) for competitive ecommerce price optimization. The proposed framework introduces a curriculum-driven Soft Actor-Critic (SAC) agent that learns sequentially against increasingly sophisticated opponents—random, naive, and rule-based. The simulation environment incorporates elasticity-aware demand, inventory dynamics, and seasonal variations to model realistic retail behavior. Across more than 9 checkpoints and multiple Monte Carlo evaluation sweeps, the SAC agent achieves up to 12.6% higher mean profit than PPO and TD3 baselines, while maintaining a fairness index (1—Gini) above 0.97 and near-zero collusion correlation. We formalize the entropy-regularized objective, provide convergence-oriented insights, and release a fully reproducible evaluation suite containing quantitative metrics, statistical significance tests, and auto-generated$\text{IAT}_{\mathrm{E}} \mathrm{X}$artifacts. Finally, we discuss governance and anti-collusion safeguards in line with current international regulatory guidance.