Analytical Learning Algorithm
An Analytical Learning Algorithm is a learning algorithm that uses prior knowledge (a domain theory) which explains the observed data to formulate a general hypothesis using deductive reasoning, and then makes sure that the hypothesis fits the observed data.
- …
- Counter-Example(s):
- See: Abductive Learning, Inductive Logic Programming, Mathematical Induction.
References
2010
- http://cse-wiki.unl.edu/wiki/index.php/Combining_Inductive_and_Analytical_Learning
- QUOTE: Two major methods of machine learning which stand in sharp contrast are those of inductive and analytical learning. Both of these methods seek to form general hypotheses which fit an observed set of data, but each method approaches the formulation of these hypotheses in a fundamentally different manner. The two approaches both have particular learning problems for which they are better suited, and they both have weaknesses which can damage the accuracies of the learners. Fortunately, several of the strengths and weaknesses characteristic to these opposing methods are complimentary, where one method fails to perform adequately the other is particularly effective. Due to this, we are able to combine the two approaches in order to take advantage of the benefits of each, forming a hybrid approach called inductive-analytical learning which has the potential to outperform its parent methods individually.
Inductive learning methods use the observed data in a training set to learn a target concept, forming a general hypothesis by finding commonalities in the data which suggest a pattern. The hypotheses learned by these methods rely on statistical inferences to make predictions and because of this they are susceptible to inaccuracy when there is not enough training data to justify their predictions. They, as with all statistical methods, are also subject to several forms of bias in the training data which can lead to erroneous predictions.
Analytical learning methods use prior knowledge which explains the observed data, referred to as a domain theory, to formulate a general hypothesis using deductive reasoning, and then makes sure that the hypothesis fits the observed data. Analytical learners can be formed with little or no training data, and so are not vulnerable to the same problem as the inductive learning methods, but they can be poorly affected by an inaccurate domain theory.
The table below from Machine Learning summarizes the contrasting attributes of the two methods presented in the above paragraphs.
- QUOTE: Two major methods of machine learning which stand in sharp contrast are those of inductive and analytical learning. Both of these methods seek to form general hypotheses which fit an observed set of data, but each method approaches the formulation of these hypotheses in a fundamentally different manner. The two approaches both have particular learning problems for which they are better suited, and they both have weaknesses which can damage the accuracies of the learners. Fortunately, several of the strengths and weaknesses characteristic to these opposing methods are complimentary, where one method fails to perform adequately the other is particularly effective. Due to this, we are able to combine the two approaches in order to take advantage of the benefits of each, forming a hybrid approach called inductive-analytical learning which has the potential to outperform its parent methods individually.
1997
- (Mitchell, 1997) ⇒ Tom M. Mitchell. (1997). “Machine Learning." McGraw-Hill.