Reinforcement learning is the mathematical formalization of trial-and-error learning: an agent learns about a system by interacting with it, receiving only scalar rewards and punishments rather than the detailed input-to-output supervision of supervised learning or the pattern-extraction setting of unsupervised learning. Its defining structural difficulties are that reward is delayed and temporally disconnected from the action that caused it, that a sequence of actions is generally required before any reward arrives, and that the environment is typically stochastic; what is learned is therefore not a single action but a policy — an association from states (perceived inputs) to actions. The framework originates in behavioral psychology's account of conditioning and reward association, and constitutes a third paradigm of machine learning alongside supervised and unsupervised learning, applicable wherever evaluative feedback exists but instructive feedback does not.
Reinforcement learning is the mathematical formalization of trial-and-error learning: an agent learns about a system by interacting with it, receiving only scalar rewards and punishments rather than …