Hi,
I understand that you are using LQG control technique for the reward function and the episode manager shows a high episode reward.
The high episode reward shown in the episode manager is indeed for the reward function. In reinforcement learning, the episode reward is a measure of how well the agent performed during a particular episode, as determined by the reward function you have defined.
In the context of LQG, the reward function might be formulated as the negative of the LQG cost, since in reinforcement learning, we typically maximize rewards, whereas in control theory, we minimize costs. The high episode reward you're observing would then correspond to a low LQG cost, suggesting that the agent's policy is effectively controlling the system with minimal error and control effort, in line with the LQG objective.
If you're using the LQG criterion for the reward function, ensure that:
- The reward function correctly reflects the LQG objectives, considering both the state and control input.
- The system dynamics are appropriate for LQG control, which assumes linearity and Gaussian noise.
- The agent's policy and learning algorithm are suitable for the problem and are effectively using the reward signal to learn.
A high reward suggests that the agent's decisions, as evaluated by your LQG-based reward function, are effectively leading to the desired outcomes. However, it is also important to evaluate the agent's performance over multiple episodes and possibly under different initial conditions to ensure consistent and robust behavior and not just a single high-reward episode.
Hope this helps!
