Information Technology Reference
In-Depth Information
the 1960's is the low tide, focusing mainly on symbolic concept acquisition.
Then in the third stage, interest in machine learning rejuvenated and many
distinctive algorithms appeared since Patrick Winston's important paper of
“Learning Structural Descriptions from Examples” in 1975. More importantly, it
was then popularly recognized that a learning system would not learn high level
concepts without background knowledge. Thus, great amount of knowledge were
introduced to learning systems as background knowledge, bringing about a new
era and new prospects for machine learning research. Due to the mass
applications of expert systems and problem solving systems, knowledge
acquisition has become the key bottleneck, to solve which heavily relies on the
advances of machine learning research. There comes the fourth stage and another
climax of machine learning research.
Main paradigms of machine learning include inductive learning, analytical
learning, discovery learning, genetic learning, connection learning, etc. (Shi
1992b). Inductive learning has been most extensively studied in the past, focused
mainly on general concept description and concept clustering, and proposed
algorithms such as the AQ algorithms, version space algorithm, and ID3
algorithm. Analogical learning analyzes similarities of the target problem with
previously known source problems, and then applies the solutions from the
source problems to the target problem. Analytical learning, e.g.
explanation-based learning, chunking, etc., learns from training examples guided
by domain knowledge. Explanation-based learning extracts general principles
from a concrete problems solving process which can be applied to other similar
problems. As learned knowledge is stored in the knowledge base, intermediate
explanations can be skipped to improve the efficiency of future problem solving.
Discovery learning is the method to discover new principles from existing
experimental data or models. In recent years, knowledge discovery in databases
(KDD, also known as data mining, DM) has attracted great research focuses,
which is considered to be a very practically useful research discipline by AI and
database researchers. KDD mainly discovers classification rules, characteristic
rules, association rules, differentiation rules, evolution rules, exceptional rules,
etc. through methods of statistical analysis, machine learning, neural networks,
multidimensional database, etc. Genetic learning based on the classic genetic
algorithm is designed to simulate biological evolution via reproduction and
variation and Darwin's natural selection paradigm. It takes each variant of a
concept as an individual of the species, and evaluates different mutations and
recombinations based on objective fitness functions, so as to select the fittest
