Hello everyone,
I am trying to develop an RL agent for a simple double integrator system. Unfortunately, my agent couldn't find the maximum reward. The attached figure is the average episode reward vs. episode number. I was wondering, what is the nature of this error and oscillator behavior of reward collection?
I would suggest to use the epsilon-greedy strategy and tune the greeedy-factor properly as function of the episode number (high exploration rate at the beginning to high exploitation, when the rewards gets better). Furthermore, adjust the learning rate over the episode number as well as the discount factor
I would suggest to use the epsilon-greedy strategy and tune the greeedy-factor properly as function of the episode number (high exploration rate at the beginning to high exploitation, when the rewards gets better). Furthermore, adjust the learning rate over the episode number as well as the discount factor
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training