Date: Tue, 14 Jan 1997 18:57:07 GMT Server: NCSA/1.5.2 Last-modified: Tue, 03 Sep 1996 18:40:11 GMT Content-type: text/html Content-length: 1637 UMass ANW Laboratory - Introduction to Reinforcement Learning

Introduction to Reinforcement Learning

In this form of learning, an agent tries to learn how to maximize a measure of long-term reward while interacting with a stochastic dynamic environment. RL is generating increasing attention in engineering, artificial intelligence, psychology, and neuroscience. It is based on the old idea that if an action is followed by a satisfactory state of affairs, or an improvement in the state of affairs, then the tendency to produce that action is strengthened, i.e., reinforced.

Receiving the most attention recently are RL problems in which the learning agent tries to maximize a measure of its long-term performance. Although similar problems have been studied intensively for many years in control engineering and operations research, the methods being developed by RL researchers have added novel elements to classical dynamic programming (DP) solution methods. Because it has direct roots in studies of animal learning, RL is also suitable for many problems faced by artificial autonomous agents in learning to interact in real-time with complex and uncertain environments.

Key featues of reinforcement learning are interactivity, uncertainty, explicit goals defined through reward functions, the problem of learning when actions have complex and delayed consequences, and the tradeoff between exploiting current knowledge and exploring to learn more.


Back to the ANW Home Page.

mcnulty@cs.umass.edu
Last Update: 11/1/94