Service-oriented architecture (SOA) is a significant framework that enables intelligent information systems to offer business process-based services, namely Business Process as a Service (BPaaS). Cloud service-based business model reconstruction within and across enterprises has become an important issue to obtain competitive advantages. Finding appropriate service component is a critical phase in enterprise collaboration, to improve the quality and correlations among collaborative service-providers. Existed methods for business processes reconstruction have not systematically and fully considered the quality correlations in finer-grain, i.e. task-level or activity-level, and temporal performance in a process model optimization. Moreover, the existed approaches might fail to work in an uncertain cloud environment where the quality parameters are unknown in advance. Q-learning has proven its worth in an uncertain cloud environment. However, advances in Q-learning are challenges to leverage in collaborative business process reconstruction. The reconstruction algorithms based on Q-learning suffered from the two core drawbacks: lack of effective exploration and extremely slow convergence property. A hybrid Evolutionary Deep Q-Learning-based BPaaS reconstruction algorithm, named as EDQL-BPR, is proposed by leveraging Particle Swarm Optimization to improve Deep Q-Learning algorithm for systematically optimizing collaborative business processes reconstruction. In order to verify the effectiveness of the proposed algorithm, an annotated transition system has been developed supporting for all possible behaviors state convention from heterogeneous initial states to the target with several collaborative information at activity level, and encoding annotated BP representation matrix automatically. Then, an autonomous three-layer framework has been built to facilitate business process discovery and service reconstruction. In this framework, our optimal collaborative services composition is treated as a multi-objective constraint optimization problem via Evolutionary Deep Q-learning with updated strategy of Q-value, and each annotated component launch a service discovery by service agent. Extensive evaluation results show that EDQL-BPR can outperform several representative business process reconstructions in terms of model optimal convergence, QoS optimality, effectiveness and efficiency under heterogeneous service selection workloads in an uncertain Cloud environment.

面向服务的体系结构（SOA）是一个重要的框架，它使智能信息系统能够提供基于业务流程的服务，即业务流程即服务（BPaaS）。企业内部和企业之间基于云服务的业务模型重构已成为获得竞争优势的重要问题。找到合适的服务组件是企业协作中的关键阶段，以提高协作服务提供者之间的质量和相关性。现有的业务流程重构方法还没有系统地，全面地考虑细粒度（即任务级别或活动级别）中的质量关联以及流程模型优化中的时间性能。此外，现有的方法可能无法在不确定质量参数的不确定云环境中工作。Q学习已在不确定的云环境中证明了其价值。但是，Q学习的进步是在协作业务流程重构中要利用的挑战。基于Q学习的重建算法存在两个核心缺陷：缺乏有效的探索和极慢的收敛性。提出了一种基于粒子群优化的混合进化基于深度Q学习的BPaaS重建算法，称为EDQL-BPR，以改进深度Q学习算法来系统地优化协作业务流程的重建。为了验证所提出算法的有效性，已经开发了带注释的过渡系统，该系统支持在活动级别具有多个协作信息的从异构初始状态到目标的所有可能的行为状态约定，并自动对带注释的BP表示矩阵进行编码。然后，建立了一个自治的三层框架来促进业务流程发现和服务重建。在此框架中，我们的最佳协作服务组合通过具有更新的Q值策略的进化深度Q学习被视为多目标约束优化问题，并且每个带注释的组件均由服务代理启动服务发现。广泛的评估结果表明，EDQL-BPR在模型最优收敛，QoS最优，