Abstract
As in other fields of Artificial Intelligence, abstraction plays a key role in learning. This chapter presents the role and impact of abstraction in two much studied paradigms of Machine Learning: Learning from examples and Learning from reinforcement. After a brief introduction to these two paradigms formulated in the KRA model, a state of the art of the use of abstraction is given for each one of them. In the former, the most widely used abstraction approaches are feature selection and feature discretization, and they are exemplified on a very simple task, and R programs are given as possible operationalization of the abstraction. The Filter, Wrapper and Embedded approaches, used for feature selection, can be extended to include many other types of abstractions, in both propositional and relational learning. Feature construction, Predicate invention, Term abstraction and propositionalization are also reviewed within the context of propositional and relational learning. In the case of Reinforcement Learning, abstraction methods can be either model driven (by analyzing the transition table and approximating it using a dynamic Bayesian network), or value driven (by analyzing the function V, and learning for it a compact representation, such as a decision tree), or policy driven. These different abstractions formulated in the KRA model support the possibility of an automatic and systematic exploration of representation changes in learning.
This is a preview of subscription content, log in via an institution to check access.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
Semi-supervised learning also learns a function that maps inputs to desired outputs, using both labeled and unlabeled examples, the latter ones providing information about the distribution of the observations.
- 2.
Guyon et al. [229] suggest to call variables the raw input variables, and features the variables constructed from the input variables. The distinction is necessary in the case of kernel methods for which features are not explicitly computed.
- 3.
There exist various R packages that support a wide variety of feature selection methods (for example the FSelector Package, which provides functions for selecting attributes from a given dataset: http://cran.r-project.org/web/packages/FSelector/index.html). Several approaches to feature selections are also available in the WEKA [565] package (http://www.cs.waikato.ac.nz/ml/weka/).
- 4.
For example, the price for sequencing one individual genome to support personalized medicine is currently still a few thousands dollars.
- 5.
Although not directly related to abstraction, active learning addresses the question of informative instances. Active learning, also called optimal experimental design in Statistics, is a form of supervised Machine Learning method, in which the learning algorithm is able to interactively make requests to obtain the desired outputs at new data points. As a consequence, the number of examples to learn a concept can often be much lower than the number required in normal supervised learning.
- 6.
There exists an R package that support a wide variety of instance selection methods, such as, for example, the “outliers” Package, which provides functions for selecting out instances from a given dataset: http://cran.r-project.org/web/packages/outliers/index.html
- 7.
There exist packages in R that support a wide variety of discretization, such as, for example, the “discretization ” Package, which provides functions for discretizing features http://cran.r-project.org/web/packages/discretization/index.html. Several approaches to feature discretization are also available in the WEKA [565] package (http://www.cs.waikato.ac.nz/ml/weka/).
- 8.
There are also scheme-driven methods, which define new intermediate predicates as combinations of known literals that match one of the schemes provided initially for useful literal combinations [504].
- 9.
To simplify the treatment, we will not explicitly represent the starting state probability distribution.
- 10.
\(\mathcal{A }\) contains the “selectors”, introduced by Michalski [368].
- 11.
As all objects have the same type, the type specification is superfluous, but we have kept it for the sake of completeness.
- 12.
An exhaustive description of decision trees is provided by Quinlan [440].
- 13.
More details can be found in Appendix G.
- 14.
In principle, also attributes for the actions could be envisaged. They can be added if needed.
- 15.
This formulation is equivalent to say that the probability is equal to \(\delta _{a,\mathtt North }\), because we are in the deterministic case.
- 16.
As mentioned before, factored MDPs exploit problem structure to represent exponentially large state spaces very compactly [76].
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
Copyright information
© 2013 Springer Science+Business Media New York
About this chapter
Cite this chapter
Saitta, L., Zucker, JD. (2013). Abstraction in Machine Learning. In: Abstraction in Artificial Intelligence and Complex Systems. Springer, New York, NY. https://doi.org/10.1007/978-1-4614-7052-6_9
Download citation
DOI: https://doi.org/10.1007/978-1-4614-7052-6_9
Published:
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4614-7051-9
Online ISBN: 978-1-4614-7052-6
eBook Packages: Computer ScienceComputer Science (R0)