Concept
Outcome Reward
A reward design method in reinforcement learning in which a reward signal is given only at the end of a task based on the final state. The signal is sparse, but the objective is clear.
Timeline (last 90 days)
Oct 26
Process reward scores intermediate steps within the multi-step trajectories an agent executes, providing a denser signal that alleviates the credit assignment problem in long-horizon tasks; outcome reward is given only at the end of a task based on the final state, with a sparse signal but a clear objective.
Unverified50%