Customer Scheduling Algorithm And Application Based On Reinforcement Learning

The customer scheduling model interacts directly with the cust,omers,which in turn can directly affect,the company s reputation,so the customer scheduling system has a very important position in a company,especially a large financial services company.A financial company has a va.riety of businesses,such as loans,repayments,balance inquiries,etc.When a customer consults with a financial company,the customer scheduling algorithm needs to guess the customer's problems based on the customer's description of the problem encountered,and assign acorrespond-ing line of business to the customer.This problem is called a dispatch task.The main goal of the dispatch algorithm is to guess the problems that the customers need to consult and to minimize the talk time between the customer scheduling model and the customers.Traditional supervised learning can only guess the problems encountered by customers,and the goal of reducing talk time cannot be quantified.Therefore,this paper proposes to use the idea of reinforcement learning to solve multi-target customer dispatch tasks.When the model assigns a large number of customers to specific lines of busi-ness,financial companies may develop a variety of consulting channels to serve customers and answer customer questions to meet the diverse needs of customers,such as chat bots,self-service apps and manual hotline systems.The task of as-signing customers to a consulting channel is called a customer diversion task.On the one hand,each channel has certain restrictions on the capacity of the cus-tomer's request.For example,the capacity of the manual hotline channel is the number of customer service personnel who can answer the customer's problems within a certain period of time.On the other hand,customers may have differ-ent preferences for different channels.Most customers tend to choose artificial hotline channels to solve the problem,but the capacity of the artificial hotline channel is relatively small.The current customer scheduling system is based on business rules and rarely considers the balance between customer satisfaction and system resources.Therefore,this paper proposes a shunting algorithm based on reinforcement learning,which can flexibly balance system resources and user sat-isfaction.When it tries to ensure that each consulting channel does not have serious congestion,it recommends the channel of high acceptance probability for customers,and increases customer satisfaction.Because the customer shunt task does not have an absolutely correct label,this paper combines three value-based reinforcement learning algorithms(double dqn.dueling dqn and prioritized expe-rience replay)to form a deep reinforcement learning model(PER-DoDDQN)to divert customers.The actual experimental data used in this paper comes from a large financial technology company.In order to effectively evaluate and train the model,we generate simulation data based on the distribution of real data,and the parameters trained on the simulation data are applied to the real data set.In order to improve customer satisfaction,this paper uses the customer's at-tribute information to infer the customer's acceptance rate of different consulting channels.Experiments show that the performance of the PER-DoDDQN-based customer scheduling model proposed in this paper outperforms the existing busi-ness rules-based customer service system.At the same time,in order to avoid the blockage of the consulting channels in the system in advance,this paper proposes to use the time series prediction algorithm to predict future customer traffic.Finally,the paper also verifies the performance of different DQN algorithms on customer scheduling tasks.Experiments show that PER-DoDDQN algorithm provides a more efficient scheduling process,which can balance system resources and customer satisfaction.