How to improve the result of Reinforcement learning (RL)?

I've recently been trying to use the reinforcement learning toolbox to train a model of a water tank that I built myself.

The model has 2-dimensional inputs and 4-dimensional states.

Reinforcement learning networks have an observational dimension of 8, including: height H(4X1), error e=Href-H(4X1).

The output dimension of action layer is 2, which is used to control the height of tanks.

The rlDDPGAgent is used for training.

The constraint of height is h1,h2=[0,280];h3,h4=[0,100],which is used for condition of isdone.

rewar function

The results of the training is too poor to control the height of tanks to Href.

I'd appreciate it if someone could answer that question.

The code and the simulink is attached.