Dynamic Service Composition Method Based on Zero-Sum Game Integrated Inverse Reinforcement Learning

Automatically generating service composition solutions that meet user application requirements is one of the hot research topics in the field of service composition in the context of Web service big data. To address the challenges of accurately obtaining reward function values and significant increase in time complexity when dealing with large-scale data in the context of reinforcement learning-based service composition, this paper proposes a novel approach that combines zero-sum game and inverse reinforcement learning. The proposed method models the service composition problem as a Markov Decision Process (MDP) and dynamically adjusts the service composition solution by solving for the optimal policy. By leveraging runtime records of service composition operations, we generate an expert experience dataset and develop a novel inverse reinforcement learning algorithm based on the integration of zero-sum game principles. Experimental results demonstrate that the proposed method effectively reduces the dependence of the inverse reinforcement learning algorithm on the quality of expert experience data and reduces the time cost of service composition.