A short description
ATRE is an ILP system that
can learn recursive theories from examples.
In particular, the problem solved
by ATRE is the following:
Given
· a set of concepts K1,
K2, ..., Kr to be learned,
· a set of objects O
described in a language LO,
· a background knowledge BK described
in a language LBK,
· a language of hypotheses LH that
defines the space of hypotheses SH
· a user's preference criterion
PC,
Find
a (possibly recursive) logical theory T, defining the concepts
K1, K2,..., Kr, such that T is complete and
consistent with respect to the set of observations and satisfies the preference
criterion PC.
Top of
this page
Experiment on the urban domain
The problem is that of
recognizing in a map downtown areas and residential areas.
Top of
this page
Experiments on the family domain
These experiments
refer to the domain of family relations used to test the MPL system. Click here to
see the genealogic tree.
Experiment 1.
The first experiment aims at learning the definitions
of ancestor, father and mother from a complete set
of positive and negative examples. In MPL the negative examples were generated
under close world assumption, and the knowledge base contained some ground
atoms concerning the predicates male, female and parent. A convenient
representation for ATRE is a single object in which all positive and negative
examples of ancestor, father and mother are explicitly reported in the head,
while all instances of male, female and parent are reported in the
body.
Experiment 2.
The second experiment of the family domain aims at
learning male_ancestor and female_ancestor from father and mother. The beam
used in this experiment is 5.
Experiment 3.
The third experiment on the family domain aims at
learning father and grandfather from an incomplete example set. The example
set contains all positive examples of father and grandfather. The negative
examples are complete for grandfather and incomplete for father, since they
are generated by means of the rule father(X,Y)=false :-
parent(X,Y)=false.
Top of
this page
Experiment on the document image understanding
domain
The problem is that of learning rules for document image
understanding. Concepts correspond to the distinct logical components to be
recognized in a document image. They are defined by different values taken by
the descriptor logic_type. The unlabelled layout objects act as counterexamples
for all the concepts to be learned, since they are instances of the concept
logic_type(X)=undefined.
Each training document is represented as an
object in ATRE, where different constants represent distinct layout components
of a page. Training examples are generated by means of the WISDOM++ system. Three long
papers, which appeared in the January 1996 issue of the IEEE Transactions on
Pattern Analysis and Machine Intelligence (PAMI), have been
considered.
Experimental results are obtained for a beam equal to 15.
In order to test the predictive accuracy of
the learned theory, we considered the fourth long article published in the same
issue of the transactions used for training.
Top of
this page
Experiment on the cognitive modeling domain
ATRE has
been applied to the problem of acquiring cognitive models in the context of the
project LHM (Learning in Humans and Machines), Task Force 1 (Representation
Change), Pilot Project 1 (Computational-Cognitive Modeling of Representational
Change in Elementary Dynamics), funded by the European Science Foundation. The
goal is that of checking whether it is possible to use machine learning systems
to reproduce the behaviour of the mental models on the development of the
concept force, proposed by cognitive psychologists and derived from
experimentation on children aged eleven or younger (Ioannides & Vosniadou,
1991). In order to build a theory corresponding to the mental models, we
provided ATRE with a set of observations extracted from a set of exercises or
situation problems that had been presented to different classes of children.
Information regarding the kinetic state of an object were obtained thorugh
observation in combination with verbal statement. The typical responses were
categorized based upon the children's interpretation. The goal is that of
learning four distinct properties of physical objects, namely when they are
stationary, when they are moving, when they have inner force and when they have
an acquired force. The conjecture that inner and acquired forces exist in
objects as their properties is not directly related to any observation and
verbal statement. Children never hear the words inner or acquired force, they
only hear the word force in different contexts. The remarkable thing is that the
same word and the same context may give rise to interpretations of inner or
acquired force depending on the prior knowledge of the observer.
Each scene
may actually describe the property of more than one physical entity, in which
case the object-centered representation adopted in ATRE proves appropriate to
describe scenes as a whole. The BK provided to the system concerns the facts
that earth and stones are instances of natural inanimate objects, while cars,
tables, sledges and balls are instances of artifacts, which are still inanimate
objects. The only animate objects specified in the background knowledge are men
and horses.
Top of
this page
Learning odd and even
The experiment aims at learning the definitions of
odd and even numbers. The problem is that of learning the concepts odd(_)=true
and even(_)=true given the predicates succ/2 and zero/1.
Top of
this page
Learning can_rich and cyclic
Top of
this page
Learning geometric figures
Top of
this page
Learning corrective rules for global layout analysis
This learning task is described in the ICML'03 paper
"Learning Logic Programs for Layout Analysis Correction". The problem is that
of learning the concepts
group(_,_)=.true,
split(_)=.vertical,
split(_)=.horizontal. Experimental results are obtained for a six-fold cross
validation.
Top of
this page
Learning tagging rules for biomedical information extraction
This learning task is described in the ILP'06 paper
"Learning Recursive Patterns for Biomedical Information Extraction". The problem is that of learning tagging rules for the biomedical literature. Concepts correspond to the distinct annotation tags to be recognized in a text. They are defined by different values taken by the descriptor annotation. The untagged tokens act as counterexamples for all the concepts to be learned, since they are instances of the concept annotation(X)=no_tag.
Training examples are generated by means of the
BEE module . Pubblications collected for the annotation of the
HmtDB genomic database have been considered. Experimental results are obtained for a six-fold
cross validation.
Top of
this page
The distribution package
There are three distribution packages available (all for Microsoft Windows platforms):
Package |
Required space on HD |
Description |
VisualATRE 2.1 (5495 KB) |
20 MB |
This package includes both the ATRE learning engine (ver. 2.3.0) and the GUI VisualATRE (ver. 2.1). ATRE has been developed in Sicstus Prolog 3.8.6 and VisualATRE has been developed in Visual C++ 6.0.
The major novelty in this release is the insertion of DATAGEN (DATA GENerator), a pre-processing component that aims at automatically generating the training and testing datasets for both the induction and the validation of logical theories. Given a training data file collecting both the training and the testing examples as well as the learning parameters, DATAGEN generates training and testing files for ATRE according to the Leave One Out, Splitting or K-Folding testing method.
Please, consult the Quick user guide of ATRE to learn more about the basic ATRE/VisualATRE usage. For a complete guide on the system, refer to on-line help provided with the system.
After downloading the file, unzip it into a temporary folder and run SETUP.EXE to install the program.
|
ATRE 2.3.0 (1086 KB) |
3.05 MB |
This package includes the last release of ATRE learning engine. The major novelty in this release is the new capability of automatically correcting the training data file in case of duplicate literals and reporting the presence of errors (variables or inconsistencies in the data). In addition, this release of ATRE is able to prevent the induction of layered theories by evaluating the training examples coverage in advance.
Download and unzip the file into a folder on HD. Run the program in line command mode.
|
ATRE 2.3.0 (no cache) (1081 KB) |
3.04 MB |
This package includes the version of ATRE without the caching capabilities. This version is less memory consuming but slower with respect to the version of system including caching techniques employment.
Download and unzip the file into a folder on HD. Run the program in line command mode.
|
Warning:
Both the systems ATRE and VisualATRE are free for evaluation, research and teaching purposes, but not for commercial
purposes.
Please Acknowledge
Top of
this page
Related publications
M. Berardi, & D. Malerba (2007). Learning Recursive Patterns for Biomedical Information Extraction. Proceedings of 16th International Conference on Inductive Logic Programming (ILP 2006), Santiago de Compostela, Spain.
M. Berardi, A. Varlaro & D. Malerba (2004). On the effect of caching in recursive theory learning. Proceedings of 14th International Conference on Inductive Logic Programming (ILP 2004), Porto, Portugal.
A. Varlaro, M. Berardi & D. Malerba (2004). Learning recursive theories with the separate-and-parallel conquer strategy. Proceedings of the Workshop on Advances in Inductive Rule Learning in conjunction with ECML/PKDD 2004, 179-193, Pisa, Italy.
A. Varlaro, M. Berardi & D. Malerba (2004). Improving efficiency of recursive theory learning. Convegno Italiano di Logica Computazionale (CILC 2004), 220-234, Parma, Italy.
D. Malerba (2003). Learning
Recursive Theories in the Normal ILP Setting, Fundamenta
Informaticae, 57, 1, 39-77.
M. Berardi, M. Ceci, F. Esposito & D. Malerba (2003). Learning Logic
Programs for Layout Analysis Correction, Proceedings of the
20th International Conference on Machine Learning (ICML2003),
27-34.
D. Malerba, F. Esposito, F.A. Lisi & O. Altamura (2001). Automated
Discovery of Dependencies Between Logical Components in Document Image
Understanding, Proceedings of the Sixth International Conference on
Document Analysis and Recognition, 174-178, IEEE Computer Society Press, Los
Vaqueros, CA.
D. Malerba, F. Esposito, A. Lanza, & F.A. Lisi (2001). Machine learning
for information extraction from topographic maps. In H. J. Miller & J. Han
(Eds.), Geographic Data Mining and Knowledge Discovery, 291-314, Taylor
and Francis, London, UK.
F. Esposito, D. Malerba, & F.A. Lisi (2000). Induction of
recursive theories in the normal ILP setting: issues and solutions, in J.
Cussens and A. Frisch (Eds.) Inductive
Logic Programming, Lecture Notes in Artificial Intelligence,
1866, 93-111, Springer, Berlin, Germany.
D. Malerba, F. Esposito, & F.A. Lisi (1998). Learning
recursive theories with ATRE, in H. Prade (Ed.), Proceedings of the
13th European Conference on Artificial Intelligence, 435-439,
John Wiley & Sons, Chichester, England.
Top of this page
Aknowledgements
Project Leader
Current Staff
Previous collaborators
Contact
Top of this page