Hi Sania,
In the context of DPG networks, while it is technically feasible to employ a U-Net architecture as both the actor and critic, it is more common and practical to convert the matrix-formatted observations and actions into vectors and utilize a standard feedforward neural network. This conversion simplifies the network design and aligns with the typical structure of DDPG implementations and b transforming the data into vector format, you can effectively leverage the capabilities of a feedforward neural network for actor-critic reinforcement learning tasks. Please let me know if you have further questions.
