Robotics Reference
In-Depth Information
before speculating on what might have happened. A knowledge base
within the program contains data from previous cases of various causes
of death, together with evidence that can either support or contradict
a particular explanation. When the investigating officers have entered
data into the program, such as eyewitness accounts, medical and forensic
evidence, Sherlock Holmes applies the knowledge from previous cases in
order to indicate the likelihood of each of a number of possible scenarios.
In this way, the program encourages the police to consider all feasible
possibilities, rather than leap to an obvious and possibly false assumption.
How Computers Learn
Humans and other animals can be observed in their gradual acquisition
of new responses to an old situation, thereby demonstrating an essential
feature of intelligence—the ability to learn from experience. Any self-
respecting robot must therefore be endowed with appropriate learning
capabilities.
The study of Machine Learning dates back to 1943, and by the time
of the Dartmouth workshop in 1956 the topic had already been em-
braced by some of the early AI researchers. Samuel's Checkers program,
19
for example, employed two different forms of learning, one of them a
simple rote learning mechanism and the other, temporal difference learn-
ing, a much more sophisticated idea that is still in use today. Here we
shall examine three additional types of learning mechanism: reinforce-
ment learning, artificial neural networks and genetic algorithms.
Reinforcement Learning
The idea behind reinforcement learning is extremely crude but rather
effective—reward success and punish failure. Employing this concept in
a computer system appears to have been first suggested by Pierre de Latil
in his groundbreaking 1953 book
La Pensee Artificielle.
20
De Latil's idea
was described using a simple game such as Tic-Tac-Toe
21
as an example.
The machine would keep track of the positions it encountered and the
moves it played during each game. If it lost a game it would discard all
19
See the section “Checkers (Draughts)” in Chapter 3.
20
Published in translation (by Y. M. Golla) in 1956 as
Thinking by Machine: A Study of
Cybernetics
.
21
Known in the U.K. as Noughts-and-Crosses.
