To improve the blocking probability (BP) performance and enhance the resource utilization, a correct decision of routing strategy which is most adaptable to the network configuration and traffic dynamics is essential for adaptive routing in optical datacenter networks (DCNs). A neural network (NN)-assisted decision-making scheme is proposed to find the optimal routing strategy in optical DCNs by predicting the BP performance for various candidate routing strategies. The features of an optical DCN architecture (i.e., the rack number N, connection degree D, spectral slot number S and optical transceiver number M) and the traffic pattern (i.e., the ratio of requests of various capacities R, and the load of arriving request) are used as the input to the NN to estimate the optimal routing strategy. A case of two-strategy decision in the transparent optical multi-hop interconnected DCN is studied. Three metrics are defined for performance evaluation, which include (a) the ratio of the load range with wrong decision over the whole load range of interest (i.e., decision error E), (b) the maximum BP loss (BPL) and (c) the resource utilization loss (UL) caused by the wrong decision. Numerical results show that the ratio of error-free cases over tested cases always surpasses 83% and the average values of E, BPL and UL are less than 3.0%, 4.0% and 1.2%, respectively, which implies the high accuracy of the proposed scheme. The results validate the feasibility of the proposed scheme which facilitates the autonomous implementation of adaptive routing in optical DCNs.

为了提高阻塞概率（BP）性能并提高资源利用率，正确决策最适合网络配置和流量动态的路由策略对于光数据中心网络（DCN）中的自适应路由至关重要。提出了一种神经网络（NN）辅助决策方案，通过预测各种候选路由策略的 BP 性能来找到光学 DCN 中的最佳路由策略。光DCN架构的特征（即机架数N、连接度D、光谱槽数S和光模块数M）和业务模式（即各种容量请求的比率R，以及到达请求的负载）作为NN的输入来估计最优路由策略。研究了透明光多跳互联DCN中的双策略决策案例。为性能评估定义了三个指标，包括（a）错误决策的负载范围在整个感兴趣的负载范围内的比率（即，决策错误E），（b）最大 BP 损失（BPL）和（c )错误决策导致的资源利用损失 ( UL )。数值结果表明，无错误案例与测试案例的比率始终超过 83%，并且E、BPL和UL的平均值分别小于 3.0%、4.0% 和 1.2%，这意味着所提出方案的高精度。结果验证了所提出方案的可行性，该方案有助于在光 DCN 中自主实施自适应路由。