I want to use the DQN algorithm from TensorFlow Agent API for a custom environment. In my environment, the agent's observations are image sequences.
As far as I know, images should be encoded by a neural network before they are used. Thereafter, the DQN network is able to produce action. I wonder how can I do this with TensorFlow Agent API.
Besides, I know that I have to override some function in order to build an environment in a custom fashion based on this tutorial.
All the examples that I have seen in TensorFlow Agent API have discrete observation space. For example, in the CartPole environment, observation space consists of two vectors describing velocity and angle.