behavioral cloning

Behavioral cloning is the most direct form of imitation: treat the demonstrations as a labeled dataset and learn the mapping from state to action with ordinary supervised learning. Each state the expert visited is an input; the action the expert chose is the label. Train a classifier or regressor to predict the expert's action, and you have a policy — no rewards, no environment interaction, no exploration.

The training objective is simply to maximize the likelihood the policy assigns to the expert's actions, equivalent to minimizing a prediction loss over the demonstration set. Because it is just supervised learning, it is fast, stable, and easy to implement, which makes it the default first thing to try and a strong baseline for any imitation problem.

Its fatal weakness is that it assumes the agent will only ever see states drawn from the expert's own distribution. The moment the cloned policy makes a small mistake it lands in an unfamiliar state, and the errors compound — the covariate-shift problem that DAgger was invented to fix.