DIAYN

Key Insight

DIAYN (Diversity Is All You Need) learns a whole repertoire of distinct behaviors with no environment reward at all — a form of unsupervised skill discovery. It hands the policy a randomly chosen "skill" code as extra input and rewards it, via an intrinsic reward, for making its visited states predictable from the skill yet different across skills — formally, for maximizing the mutual information between the skill code and the states it produces. The effect is that each skill claims its own region of behavior — one learns to run forward, another to flip, another to stand still — without anyone hand-designing those goals. Why it matters: these pre-trained skills become a ready-made action vocabulary that a later reward-driven task can compose, and the idea connects directly to maximum-entropy RL, which similarly prizes a diverse, non-collapsed policy.