reward-to-go (partial return)

Reward-to-go is the part of the return that still lies ahead of you. Standing at time step t, you do not care about rewards you already collected — they are sunk, fixed, gone. What matters for the decision you make now is everything you can still earn from this point onward: the partial, remaining return. It is the same quantity as the full return, just measured from a later starting line.

It is the tail sum of the discounted rewards from step t onward. This idea sharpens policy-gradient algorithms: when crediting an action for an outcome, only the rewards that came after it should count, since an action cannot influence the past. Using reward-to-go instead of the whole-episode return removes irrelevant earlier rewards from each action's signal, which lowers variance and speeds learning without introducing any bias.