The Design And Implementation Of Decision Optimization System Based On C-MDP

In recent years, with the rapid development of Telco, Banking, Retail and others, internal competition in every of these industries is more and more intense. CRM departments of operators in these industries manually conduct promotional campaigns for the purpose of customer retention, cross sell and up sell. By this way the revenue had been improved in short time, however these methods do not consider the uncertainty in campaigns, which can’t maximize operators’revenue. Decision optimization, which use algorithm to analysis historical data and modeling, provides the best marketing decision and recommend to operators to solve the problems above.C-MDP, which is Constrained Markov Decision Process, is an algorithm applicable for decision optimization system. Compared to marketing decision made manually, C-MDP figures out the best decision for every group taking constraints into consideration.This paper introduces a Decision Optimization called NBAOPT, which is divided into Action Cluster and NBAOPT Studio. Action Cluster performs Binning and Clustering processes on users’historical data to classify customers into different groups using Python. The output of Action Cluster is the input of NBAOPT Studio. NBAOPT Studio defines the set of entity, state, action, policy, resource and constraint of C-MDP in XML format. The algorithm of computing reward during state transition in C-MDP is implemented by Java. The algorithm computes the optimized reward by iteration to figure out the best decision and then outputs the models. At last, implement a visualization utility using Dojo and Apache Wink framework to analysis and evaluate models.