Deep reinforcement learning-driven smart and dynamic mass personalization

Smart mass personalization is becoming increasingly important to improve the competitiveness of products. In mass personalization, customers' contextual data is characterized by complexity and fluctuation. Hence, designers must ensure the timeliness of smart mass personalization that can continuously satisfy customers' demands. This paper proposes a deep reinforcement learning-driven system for dynamic and smart mass personalization. Besides, the system adopts deep Q-network as the training algorithm due to its compatibility with both off-policy and on-policy training. In the beginning, deep Q-network will get trained based on previous customers' contextual data collected from purchase history and web services until it can generate the expected policy for concept generation. Then, the agent in deep Q-network will dynamically tune the algorithm by continuously interacting with incoming customers' contextual data. Besides, this paper depicts a scenario of personalization for automobiles to illustrate this system. The contribution of this paper lies in the application of DRL to realize dynamic updates in smart mass personalization and the innovative dynamic action space generated from customer clusters.