A Multi-agent Reinforcement Learning Based CR Allocation Approach For Multi-Scenario Advertising Systems

The goal of a recommender system is to recommend the most suitable items to a user from a large number of candidate items. The computational cost increases with the number of user requests and the complexity of the service or model. Under limited computing resource, how to make a tradeoff between computing cost and business benefit has become an important issue. This study focus on computation allocation (CR) optimization in multi-scenario advertising systems. A well designed multi-step MDP model is introduced, with each agent managing CR control task in their respective scenario. The MAPPO-CR algorithm is developed to enhance convergence and improve allocation performance through tailored differential rewards and cooperative mechanism. Experiments demonstrate significant improvements in convergence speed, allocation revenue and scalability, highlighting the effectiveness of the proposed allocation algorithm for complex dynamically changing online environments.