Autonomic optical transmission and networking requires machine learning (ML) models to be trained with large datasets. However, the availability of enough real data to produce accurate ML models is rarely ensured since new optical equipment and techniques are continuously being deployed in the network. One option is to generate data from simulations and lab experiments, but such data could not cover the whole features space and would translate into inaccuracies in the ML models. In this paper, we propose an ML-based algorithm life cycle to facilitate ML deployment in real operator networks. The dataset for ML training can be initially populated based on the results from simulations and lab experiments. Once ML models are generated, ML retraining can be performed after inaccuracies are detected to improve their precision. Illustrative numerical results show the benefits of the proposed learning cycle for general use cases. In addition, two specific use cases are proposed and demonstrated that implement different learning strategies: (i) a two-phase strategy performing out-of-field training using data from simulations and lab experiments with generic equipment, followed by an in-field adaptation to support heterogeneous equipment (the accuracy of this strategy is shown for a use case of failure detection and identification), and (ii) in-field retraining, where ML models are retrained after detecting model inaccuracies. Different approaches are analyzed and evaluated for a use case of autonomic transmission, where results show the significant benefits of collective learning.