Curriculum Reinforcement Learning
I first encountered this idea through this paper about Head-to-Head Autonomous Racing
Curriculum Reinforcement Learning (CRL) is a method that uses a curriculum, which is a sequence of tasks that gradually increase in difficulty, to train policies.
The idea behind CRL is to start with simple tasks that the robot can easily solve and then gradually increase the complexity of the tasks as the robot becomes more skilled. This allows the robot to learn the task incrementally, which can improve the sample efficiency and the final performance.
Links
Paper: https://rpg.ifi.uzh.ch/docs/ICRA21_Yunlong.pdf
3 stages:
- Racing (pure pursuit)
- Overtaking
- Avoidance
These stages build on top of each other. The reason they do this is to prevent training from getting stuck in a local minimum.