Logical view. This is the last view that will be developed, given that the technical view would refer to the connection to the embedded software or hardware of the physical devices itself, that is out of scope for this document. Commented [M(130]: Is this true? Commented [A(131R130]: Should be value-based instead of DQN, I fixed it In the logical view the detail of the chosen implementation and how the policies are updated per RL algorithm family will be highlighted. The two families are namely value-based and actor-critic. At this level, there are implementation specific choices that are unavoidable at this level of depth. From the AI perspective, it is interesting to understand what we called the ‘Predict’ block and to understand how the policy is updated in RL, what we called the ‘Update policy’ block. Those are the AI specific elements of the flows. For the EM usecase, applying a DQN algorithm (value-based), the ‘Predict’ block is implemented as follows: Version Status Date Page 2.0 Non-Confidential 2024.05.1172022.03.1 87/100 Figure 73 From the perspective of the policy update a vanilla DQN implementation was used, also from the library baselines3. DQN is a value-based model free RL algorithm that works according to working in the graph below. The Q-network corresponds to the custom-defined network defined above. The Target Q-network in a copy of the network that is only updated every policy update times, as can be seen in graph x in the functional view. Figure 74 In the UUV case which implements an actor-critic algorithm (see chapter 4), the Predict block contains a neural network called “the actor” with the following dimensions:
Logical view. As previously discussed at a logical level, the platform will offer different types of services. These services can be conceptually divided into three layers: content layer, mapping layer and user layer.12 Content Layer 11 xxxx://xxxxxxxxxxxxx.xx/patterns/data/saga.html 12 as stated in the CPN Grant Agreement (Part B p.13-14) The Content Layer will focus on the extraction of relevant information from the different content sources or content providers. This layer is composed of two types of services: (1) content procurement (for structured and unstructured heterogeneous data streams gathering) and (2) knowledge extraction (for user personalisation such as relation identification and clustering)
Logical view. The diagram below depicts the overall logical reference architecture of the USP and the related logical components.
Logical view. The Remote Maintenance System consists of a web-application, mobile Apps and a Smart Glasses application (see Figure 5-10). The main functionalities are the management and request of technical documentation, a live-expert remote service, machine data acquisition (not used in Prophesy), a knowledge database system and management functionalities.
