transfer in RL

Transfer in RL is the everyday idea that what you learned on one task should give you a head start on the next. Someone who plays tennis picks up squash faster than a total beginner; an agent that has learned to walk on flat ground should not have to relearn everything to walk on a slope. Formally, transfer reuses knowledge — a policy, a value function, learned features, or a world model — gained on a source task to speed up learning on a related target task.

It comes in several flavors. You can transfer representations (the convolutional features that see the game), warm-start a policy and fine-tune it, distill several teachers into one student, or transfer a dynamics model that is largely task-independent. The benefit is measured as jumpstart (better initial performance), faster learning, or a higher final score. The constant danger is negative transfer: when source and target differ in the wrong way, prior knowledge actively hurts and the agent would have been better off starting fresh.

Transfer is foundational to nearly everything else on this frontier — meta-RL, continual learning, and sim-to-real are all transfer with extra structure. The open question is how to transfer reliably without hand-picking what to carry over: we still lack a principled way to know in advance which pieces of knowledge will help a new task and which will mislead it.