Towards Transfer Learning for Revenue and Pricing Management

Reinforcement Learning (RL) has proven itself as a powerful tool to optimize pricing processes. With the support of deep non-linear function approximation tools, it can handle complex and continuous state and action spaces. This ability can leverage the utility of pricing algorithms in markets with a vast number of participants or in use cases where additional product features should be considered in the pricing system. One problem with those tools is their apparent demand for training data, which might not be available for a single market. We propose to use techniques instead, that leverage the knowledge of different problems. Several similar algorithms have been proposed in the past years to allow RL algorithms to operate efficiently on various processes simultaneously. DISTRAL continuously merges information from different decision processes towards a distilled policy and uses the joint policy to update the market-specific source policies. We will discuss the influence of such regularization mechanisms. Multi-market pricing problems are used to illustrate their impact.