Abstract:

A randomized reinforcement learning approach for Web Service (WS) composition is proposed to determine the optimal WS to execute the task in a predefined process model. This approach can learn the optimal task allocation policy corresponding to the task in a given model based on actual performance of WS. It can also identify the optimal WS corresponding to the task in a given model based on task allocation policy. By this approach, the coordinated execution can continually satisfies user's service request to the optimal levels. Compared with other similar methods, the approach integrates the exploitation of acquired data about the past performance of individual services with the optimal, undirected, continual exploration of new composition according to QoS actual performance so as to learn the optimal WS composition corresponding to the process method. The relation between exploitation and exploration is accounted for discussing the range of the entropy.

Key words: computer application, composite service, predefined process, reinforcement learning, service choice, composer