Information Technology Reference
In-Depth Information
Table 12.1
The four features (Features type 1) associated with a number sequence
Feature
Meaning, Example
F1
The sequence is either monotonically increasing or decreasing,
1234567
F2
All the numbers in the sequence have the same parity,
2468
F3
The parity of the numbers in the sequence alternate,
1234567
F4
The size of each number in the sequence is either monotonically increasing or
decreasing,
1 10 100 1000 10000
this case, a numerical sequence. The concepts are the target classes of a Bayesian
learning mechanism for which this specific set of features provides information. The
four features of a sequence are shown in Table
12.1
below.
Information about the prediction of a hypothesis is deduced and provides feedback
to the retroduction process to refine and improve the hypothesis if there is a need for
it. This result also serves as a framework for inductive testing. The consequence of
induction causes further retroduction. The interaction of the three types of inference
suggests why, while it is possible to generate a large number of hypotheses for given
observable facts, people tend to generate and accept the first valid hypothesis that
satisfies their preconceived criteria (Wason and Johnson-Laird
1968
).
12.2
Implementation of Intelligence
The role of the controller, as shown in Fig.
12.1
, is to monitor the progress of
the creation, the prediction and the validation of hypotheses. It also keeps track of
whether a hypothesis was successfully generated and validated during each cycle.
The ovals describe the main mechanism, the clear boxes indicate ordered lists
of concepts, etc., and the shaded boxes give the information and its structure to be
processed by the main mechanism. The 'thick' arrows pick out the main processing
cycle.
The controller governs the process that assesses the relative confidence for the
hypothesis generated by each concept. In one set of tests extending the original
simple Bayesian learning into a 'running' Bayesian learning system generates the
confidence. The controller feeds a portion of the number sequence to the retroduction
by taking a sample of the extended sequence. This is achieved by partitioning the
sequence into “windows” and each “window” will be subjected to retroduction at
each cycle. Figure
12.2
shows two methods (A and B) of using a running window.
A window (w
x
) is a 'running' sample of a sequence. If X
1
to X
m
is a sequence, then
“windows” for X (size
n
) are generated as:
