Reinforcement Learning
/ˌriːɪnˈfɔːsmənt ˈlɜːnɪŋ/re·in·force·ment learn·ingnoun
Definition
1.[in machine learning] a training paradigm in which an agent improves its policy purely from rewards collected while acting in an environment, with no labelled examples of correct behaviour to imitate.
Reinforcement learning taught the arm to grasp without anyone demonstrating the motion.
Formal statement
Q(s,a) <- Q(s,a) + alpha [ r + gamma max_a' Q(s',a') - Q(s,a) ]The Q-learning update. gamma discounts future reward; alpha is the step size.
Etymology
From reinforcement, the term Pavlov's translators used for a stimulus that strengthens a conditioned response, itself from Latin re- 'again' plus fortis 'strong'. Carried into machine learning from animal-behaviour psychology by Minsky in the 1950s.
Synonyms
- trial-and-error learningsense 1 · Near
Antonyms
- supervised learningsense 1 · Relational
See also
References
- Sutton, R. & Barto, A. (2018). Reinforcement Learning: An Introduction, 2nd ed.MIT Press.
- Watkins, C. & Dayan, P. (1992). Q-learning.Machine Learning, 8, 279–292.