Artificial Intelligence for Smarter Power Systems by Simões Marcelo Godoy;

Author:Simões, Marcelo Godoy;
Language: eng
Format: epub
Publisher: Institution of Engineering & Technology
Published: 2021-07-04T16:00:00+00:00

The development of NNs was inspired by the studies for understanding the biological nervous system. Preliminary theoretical foundations on physiology and psychology for neural networks were proposed by Alexander Bain (1873) and William James (1890). In their work, both thoughts and body activity resulted from interactions among neurons within the brain. Their concepts foretold the notions of a neuron’s activity as being a function of the sum of its inputs. Half a century later, McCulloch and Pitts (1990) published a seminal paper, in which they derived theorems related to models of neuronal systems based on what was known about biological structures in the early 1940s, showing that a network could represent any finite logical expression with a massively parallel architecture. In 1949, Hebb (1949) published a book, where he defined a method to update synaptic weights for what is now referred to as Hebbian learning. The landmark by Rosenblatt (1962) defined an NN structure called perceptron. It was simulated in detail on an IBM 704 computer at the Cornell Aeronautical Laboratory and caught the attention of engineers and physicists because such a computer-oriented paper described the perceptron as a “learning machine.” This paper laid the groundwork for both supervised and unsupervised training algorithms as they are today in backpropagation and Kohonen networks, respectively.

In 1960, Widrow and Hoff (1960) published a paper where they had simulated NNs in computers and also had implemented their designs in hardware. They introduced a device called an ADALINE, an adaptive linear processing unit based on a neuron. An ADALINE consists of a single neurode with an arbitrary number of input elements that can take on values of plus or minus one and a bias element. Before being summed by the neuron-summer circuit, each input (including the bias) is modified by a gain. The Widrow–Hoff algorithm is a form of supervised learning that adjusts the weights according to the error intensity at the output of the summer. They have shown that their technique for adjusting the weights could minimize the sum-squared error over all patterns in the training set.

The first wave of artificial NNs (ANNs) research can be associated with the frequency of the keyword “cybernetics,” which started around 1940s and peaked at 1970s. The development of NNs slowed at the end of the 1960s and middle of the 1970s, mainly because Minsky and Papert (1969) cooled off the NN community with a book called Perceptrons. In such a book it was presented a comprehensive analysis on a single perceptron, without hidden layers. However, their writing style was full of criticism in claiming that most of the research about NNs was “without scientific value.” They showed that the two-layer perceptron was rather limited, because it could only work with problems associated with linear separable solution spaces. The Exclusive-OR (XOR) problem was used on an elementary system that the perceptron was unable to solve. About the year 1969 people only knew how to train two-layer networks; there was no effective algorithm to train a network with three or more layers.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.