curriculum learning (RL)

Just as you learn arithmetic before calculus, an RL agent learns faster when its tasks are ordered from easy to hard. Thrown straight at a brutally hard task it may never receive any reward at all and never get off the ground, whereas a curriculum hands it early wins to build on.

You arrange a sequence of tasks or environment settings of increasing difficulty — shorter distances, slower opponents, fewer distractors — and promote the agent as it masters each stage. Difficulty can be hand-designed, generated automatically (the environment proposes tasks just beyond current ability), or driven by self-play, where the opponent is always a near-equal. Goal generation and reverse curricula (start near the goal and push the start farther away) are popular variants.

A poorly chosen curriculum can mislead: skills learned on easy versions may fail to transfer, or the agent may overfit to early stages. The ordering has to lead genuinely toward the final task, not merely toward whatever is locally easiest.