Learning agent for operating in ML control device and method for adjusting cost function of learning agent

There is provided a method for adjusting a cost function of a learning agent, said learning agent being arranged for operating in a Machine Learning, ML, based control device for controlling a system, wherein said learning agent is associated with state spaces and action spaces which define possible states of the system that the learning agent may receive from the system and possible actions that are available to the learning agent for applying to the system, respectively, wherein said method comprises the steps of: applying, by said learning agent, said cost function to received states for mapping to actions suitable for applying to the system; receiving, by said learning agent, from an expert agent, actions, wherein said actions are based on the states that are also received by said learning agent; adjusting, by said learning agent, said cost function based on said received actions from said expert agent.