Hands-On ROS for Robotics Programming

Hyperparameters and optimization

So, how can you choose the values of alpha, gamma, and epsilon? Well, this strategy has to be based on both intuition and trial and error. In any case, the three of them should decrease over time as the agent learns the best actions:

• Decreasing the need for learning, alpha, as the agent knows more about the environment and may trust in the acquired experience
• Also decreasing the discount factor, gamma, as an agent develops an end-to-end strategy and not only focuses on the immediate reward
• And, finally, decreasing the exploitation rate, epsilon, because the exploration gains lose priority as the environment is well known

At this point, you should be ready to enter OpenAI ROS to train agents that power robots ...