Author: Marcus Hutter (1990)
Superviser: Gerhard Weiß
Reference: TU-München,
Fortgeschrittenen-Praktikum
Department: Theoretische
Informatik und Grundlagen der KI
Subj-class: Artificial Intelligence; Learning
ACM-class: I.2; I.2.6; I.2.8; F.1.3
Comments: 30 pages in German with Listing
Keywords: Neural nets, reinforcement/unsupervised/supervised learning,
Hebb nets, Hebb learning rule, XNOR.
Abstract:
This Fopra is motivated by the following observations about human
learning and about human neural information processing. On the one
side humans are able to learn supervised, unsupervised and by
reinforcement, on the other side there is no neural distinction
between informative, uninformative and evaluative feedback.
Furthermore, the Hebb learning rule is the only biological
inspired learning mechanism. If the human brain is indeed a Hebb
net this would imply that Hebb nets are able to learn by
reinforcement. The goal of this Fopra is to investigate whether
and how Hebb nets could be used for reinforcement learning. It
is shown that Hebb nets with a suitable prior net topology
can indeed learn, at least simple tasks, by reinforcement.