Challenge 2: Rewards

How would you reward an agent to encourage good behavior? We ultimately want the hummingbird to stick its beak into a flower and keep it there for a short time. Would you reward it for getting close to the flower? What about hitting the flower from behind? What if the bird touches the flower with its butt?

ML-Agents will take random actions until they find which actions produce the highest reward. If the task is too hard, they might fail to ever find the best actions. If the rewards are not well designed or are unbalanced, they might find a way to cheat to get the most points.