Hi Aysegul,
I understand that you are facing some confusion in incorporating hard constraints in your reinforcement learning system. Usually, the decision to stop the training after a violation of hard constraints depends on the type of problem. In your case as you are working on a physical system, violating hard constraints can cause incorrect, illogical, or non-sensical results. So, it is better to stop the training when the system violates a hard constraint.
There exist better ways to incorporate hard constraints in the reinforcement learning environment such that the learner is incentivized to stay within the constraints. Some of the methods are listed below:
1) Reward Shaping:
You can shape the reward function to provide explicit penalties whenever the agent violates a constraint or crosses a boundary. By assigning negative rewards for constraint violations, the agent learns to avoid such actions.
2) Barrier Methods:
Barrier methods involve introducing a barrier function that penalizes constraint violations. The barrier function increases as the agent approaches the constraint boundary, discouraging violations.
3) Constraint Enforcement:
You can also use the constraint enforcement block in Simulink to model your hard constraints directly into the physical system so that the model never violates the constraints.
You can find a detailed example of constraint enforcement on the link mentioned below:
I hope this helps.
