17-03-2014, 09:54 PM
Introduction And Background
An inductive learning system is usually considered to have
attained a concept if it is able to distinguish objects that
embody the concept from those that do not. However,
Simon and Kotovsky pointed out that there are certain kinds
of concepts which we commonly measure attainment by the
ability of a system to produce an object satisfying the concept
rather than the mere ability to identify an object as belonging
to that concept [14]. For example, a system is considered to
have acquired the concept ‘simple alternation of a and b’
embodied in the sequence ‘abnbabababa-’ only when it is
able to extrapolate the sequence by producing the next characters
(i.e. ba) that satisfy the concept.
Since the ability of a learning system to acquire the concept
underlying an ordered sequence of objects permits it to
predict future objects based on the acquired knowledge, the
task that such a system performs may be referred to as prediction.
We have identified two different types of prediction
problems: the deterministic prediction (DP) problem and the
probabilistic prediction (PP) problem. Predicting succeeding
letters in a letter sequence such as abababababa- is deterministic
since they are completely determined by the preceding
ones. Predicting weather conditions based on past
records is, however, probabilistic since weather forecasts cannot
usually be made with complete certainty. In this paper,
an inductive learning algorithm that is able to handle these
problems is presented.
An inductive learning system is usually considered to have
attained a concept if it is able to distinguish objects that
embody the concept from those that do not. However,
Simon and Kotovsky pointed out that there are certain kinds
of concepts which we commonly measure attainment by the
ability of a system to produce an object satisfying the concept
rather than the mere ability to identify an object as belonging
to that concept [14]. For example, a system is considered to
have acquired the concept ‘simple alternation of a and b’
embodied in the sequence ‘abnbabababa-’ only when it is
able to extrapolate the sequence by producing the next characters
(i.e. ba) that satisfy the concept.
Since the ability of a learning system to acquire the concept
underlying an ordered sequence of objects permits it to
predict future objects based on the acquired knowledge, the
task that such a system performs may be referred to as prediction.
We have identified two different types of prediction
problems: the deterministic prediction (DP) problem and the
probabilistic prediction (PP) problem. Predicting succeeding
letters in a letter sequence such as abababababa- is deterministic
since they are completely determined by the preceding
ones. Predicting weather conditions based on past
records is, however, probabilistic since weather forecasts cannot
usually be made with complete certainty. In this paper,
an inductive learning algorithm that is able to handle these
problems is presented.