Where does reinforcement collide?

A consequence that follows a behaviour and makes that behaviour more likely in the future. B. F. Skinner set out the principle in The Behavior of Organisms (1938) and Science and Human Behavior (1953): a reinforcer is defined only by its effect, by whether the response it follows grows more frequent, and not by whether it feels pleasant. Positive reinforcement adds a stimulus, negative reinforcement removes an aversive one, and both raise the behaviour; neither is punishment, which lowers it. The referent is an observable response and the unit is a change in its rate. This meaning names a relation between an act and its consequence, not an algorithm and not a physical material.[1][2]

Unit

A consequence's effect on a behaviour's future probability

Recovery means

Not applicable

Measured by

Change in the rate or probability of a response following the consequence

A family of algorithms in which an agent learns a policy by interacting with an environment and adjusting its actions to maximise a scalar reward. Sutton and Barto give the standard account in Reinforcement Learning: An Introduction (2018): the reward is a signal defined in advance by the designer, and the agent's problem is to find actions that return the most reward over time. Mnih and colleagues (2015) trained such an agent to play Atari games directly from pixels. Here reinforcement is a computational method whose referent is a policy over states and actions, and whose unit is expected cumulative reward. Unlike the behaviorist meaning, the reward is stipulated first and the behaviour is learned to fit it, rather than the reinforcer being discovered from whether behaviour rose.[3][4]

Unit

A reward signal guiding an agent's learned policy

Recovery means

Not applicable

Measured by

Cumulative reward, the return, that the agent's policy achieves over interaction