Dynamic Pricing on E-commerce Platform with Deep Reinforcement Learning
In this paper we develop an approach based on deep reinforcement learning (DRL) to address dynamic pricing problem on E-commerce platform. We models real-world E-commerce dynamic pricing problem as Markov Decision Process. Environment state are defined with four groups of different business data. We make several main improvements on the state-of-the-art DRL-based dynamic pricing approaches: 1. We first extend the application of dynamic pricing to a continuous pricing action space. 2. We solve the unknown demand function problem by designing different reward functions. 3. The cold-start problem is addressed by introducing pre-training and evaluation using the historical sales data. Field experiments are designed and conducted on real-world E-commerce platform, pricing thousands of SKUs of products lasting for months. The experiment results shows that, on E-commerce platform, the difference of the revenue conversion rates (DRCR) is a more suitable reward function than the revenue only, which is different from the conclusion from previous researches. Meanwhile, the proposed continuous action model performs better than the discrete one.
PDF Abstract
