Robotics Reference
In-Depth Information
It is likely, however, that my judgement was sometimes impaired
by fatigue. [3]
ArtificialNeuralNetworks
The neurons in the human brain are rather simple processors, each of
which receives input signals from many other neurons, then merges the
information from these input signals, and finally sends out many signals
to other neurons based on this merged information. It is the network of
these neurons, the huge number of neurons and their connections, that
create the immense processing power of the human brain.
Artificial Neural Networks (ANNs), often called simply “neural net-
works” or “neural nets”, are a computer-based representation of how psy-
chologists and neurologists believe this process works in our brains. The
first artificial neurons were devised in 1940 by Warren McCulloch and
Walter Pitts, who were able to prove that, by linking together several
simple processing units (artificial neurons) to form a network, it is possi-
ble to create in the whole network more computational power than one
would expect from the sum of its units. Each unit simulates a biological
neuron, receiving one or more inputs and producing an output based on
the combined information from these inputs.
In 1957 Frank Rosenblatt, a psychologist working at the Cornell
Aeronautical Laboratory, invented the “perceptron”, which consists of
a layer, or possibly more than one layer, of artificial neurons. A neuron
receives one or more input signals and assigns a numerical weighting to
each signal, akin to the numerical weightings employed in the evaluation
functions in game playing programs.
24
By multiplying the strength of
each input signal by its corresponding weighting and calculating the to-
tal of these products, an artificial neuron is able to compute a numerical
score, akin to an evaluation function's score for a position in a game. If
this score for an artificial neuron is above a certain threshold, the neuron
is said to “fire”.
Just as perceptrons are made up of one or more layers of artificial
neurons, so ANNs are made up of a network of perceptrons and were
originally called multi-layer perceptions. The more artificial neurons and
layers there are in a network, the more connections there will be between
them and the more knowledge an ANN will be able to acquire, since it
is the connections that hold the knowledge.
24
See the section “How Computers Evaluate a Chess Position” in Chapter 3.
