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