AI subsystem – operational requirements. In the operational phase, as an extension to the training phase when successful, all the requirements that needed to be approximate first, need now to be more definite. Version Status Date Page 2.0 Non-Confidential 2024.05.1172022.03.1 84/100 For deployment upon a real system, the hard requirements need to be specified and the solution needs to be finetuned in order to function as described in all aspects of the requirements. The solution in training phase and operational phase should be the same, as varying the conditions may result in needing another iteration on the training phase. The AI software is known to work very well for a very specific range of conditions, so varying them may result in another iteration. The extent of the changes in data or requirements will determine the extent of the rework in the training phase. In general, if the changes are small, the method might be correct, but the models will need to be retrained, possibly needing other processing, hyperparameters or length of training. What makes a small or big change is up to the experts to decide, when the changes of requirements are too large not even the approach can be assumed to be right anymore. So, having changes in the requirements is especially critical for AI systems, that need a good definition of the problem beforehand. In training those requirements need be less accurate, but in big lines, correct. Assuring this for a smooth stepping to deployment is the key. For the electron microscope the first months there was no specific business-case (but a general problem statement) and a more generic approach to the problem was taken. When the business case was ready the approach was still holding even though the hyperparameter tuning was the key to solve the problem. The same also holds for UUV Use Case. In the operational phase, it is important to monitor the performance of the RL-Model in hindsight whether it achieves the intended goal of creating critical scenarios. Here, fully documentation of the processes and the possibility of user interacting with the resulting data is important for the users to evaluate and adjust the model for possible changes. Both the model-logging infrastructure (Figure 69) and the data management (Figure 70) required in the training of the model is helpful to fulfill those tasks.
