Immediate feedback on our actions helps us to learn. However in real life we may have intermittent rewards, only occasionally having some form of benefit or cost which may be based on long past actions, for example, feeling backache the morning after digging the garden. This is a major issue for reinforcement learning in robotics and agent based systems, which either need to trace back from a reward to the actions that were its ultimate cause, or create a predictive model of future rewards.
Defined on page 379
Used on Chap. 16: page 379
Reinforcement learning with intermittent rewards.