imitation learning

Imitation learning is teaching by showing instead of by reward. Rather than letting an agent stumble through millions of trials to discover what works, you hand it a set of expert demonstrations — recordings of a skilled human or controller performing the task well — and ask it to reproduce that behavior. A student watching a driving instructor steer through traffic and then copying the moves is doing imitation learning.

Concretely, a demonstration is a trajectory of states usually paired with the actions the expert took. The learner fits a policy that maps states to actions, either by plain supervised copying (behavioral cloning) or by first recovering the expert's hidden goal (inverse RL). This sidesteps two hard parts of standard RL: hand-designing a reward function and exploring safely, both of which the expert has implicitly solved for you.

The catch is that you only learn what was shown. The agent has no model of why the expert acted, so it can fail badly in situations the demonstrations never covered, and it rarely surpasses the demonstrator unless an extra reward signal is added.