<![CDATA[Dynamic pricing is crucial for e-commerce platforms to adapt to rapidly fluctuating market demands. However, conventional reinforcement learning methods, such as Tabular Q-Learning, suffer from the curse of dimensionality and discretization errors. Furthermore, the Deep Q-Networks (DQN) standard often encounters overestimation bias, leading to aggressive price drops and severe financial losses known as the Death Spiral. This study aims to address these issues by proposing a continuous Markov Decision Process (MDP) environment using a Double Deep Q-Network (DDQN) architecture. The model is trained using historical time-series data from the Brazilian E-Commerce Public Dataset by Olist. The proposed innovations include internal State Normalization to eliminate Gradient Domination, and adaptive Reward Shaping paired with absolute economic Guardrails to strictly prevent the AI agent from setting prices below the base cost. Comparative evaluation results demonstrate that the DDQN agent achieved a highly stable convergence, generating a maximum cumulative profit of IDR 11,146,669. This proposed model significantly outperformed the static pricing baseline (IDR 3,975,383) and the conventional Q-Learning algorithm (IDR 9,966,507). These findings prove the robustness of the proposed DDQN architecture in safely, precisely, and automatically exploiting continuous market inelasticity without relying on manual interventions.]]>
Copyrights © 2026
