I am training a RL agent whose state is composed of two numbers, ranging between 4 ~ 16 and 0 ~ 360. The action is continuous and between 0~90. In real life, the states can be any I am training a TD3 agent using the stable baselines library. In real life, the state may be any pair of numbers in the aforementioned range. Hence, I am generating random numbers for training. Leading to different data at each episode. I have realized that the trained agent is predicting actions just in the boundaries of the action range. Could this issue be caused by using different data for the different episodes of the training?.
In the real application of my algorithm the pair of numbers will be arbitrary, given that they are in the corresponding range. By random numbers for training I mean that I am generating pairs of numbers using uniform distributions between the given boundaries. And using them to train the system. The reward is a function of these numbers, however it does not have an analytic expression.