When should I use Reinforcement Learning instead of supervised learning?

All answers to this question.

Reinforcement Learning is specifically designed for "Sequential Decision Making" in dynamic environments where there isn't a single "correct" label for every step. If your robot needs to adapt to moving obstacles or changing priorities, RL is the way to go. In supervised learning, you need a massive labeled dataset of "perfect" examples, which is hard to get for every possible warehouse configuration. With RL, you define a reward function—like $+10$ for reaching the goal and $-5$ for a collision—and let the agent explore. Use a simulation environment like OpenAI Gym or NVIDIA Isaac Sim to train the agent before putting it on physical hardware.

   Answered 2024-09-22 by Patricia Young

Defining a reward function is notoriously tricky; have you considered starting with "Imitation Learning" where the agent first learns from human demonstrations before fine-tuning with RL?

   Answered 2024-09-28 by Richard Taylor

• Richard, I hadn't considered Imitation Learning! That would solve the "cold start" problem where the agent just wanders aimlessly for the first few thousand episodes. If I can record a few hours of manual robot operation, that could give the agent a baseline "policy" to start from. Then, I can use RL to optimize that policy for speed and energy efficiency. This seems much safer than letting a completely untrained agent loose in a virtual simulation where it might learn weird, non-physical exploits.

     Commented 2024-10-04 by Christopher Lee

RL is powerful but expensive. If the environment doesn't change that much, sometimes a simple A* search algorithm with some light ML for obstacle prediction is more reliable.

   Answered 2024-10-12 by Nancy Davis

• Nancy makes a great point. Always ask if you actually need a "brain" or if a sophisticated "map" will do the job with less overhead.

     Commented 2024-10-18 by Patricia Young