The technology of mashine learning for a composite web service development

We analyze dynamic programming and machine learning algorithms (on example of Q-learning) used for automatic adaptive composition of web services based on service quality assessments, their input parameters and work specifics. Software implementation of these algorithms on sets of services of different volumes is developed for comparison their performance parameters. We determine that the considered methods allow finding the optimal set of services only for composition with a predefined fixed-length route. This restriction causes a need to generalize the problem formulation for an arbitrary set of service classes in the composition route. On base of the performed analysis, we developed an algorithm that solves this problem of building a composite service with a route of arbitrary length (using the Q-Learning method), that has the best overall quality ratings. A software implementation of both this algorithm and other algorithms for solving this problem (genetic algorithm, greedy search, dynamic programming, SARSA, etc.) are developed to compare the speed of their work and the evaluation of the resulting composite service on data sets of different volumes.