Special Topics:

Evolutionary Construction of Decision Trees

339.339

1KV

Sprogar

Block

Begin: 16.10.2013

Decision trees are transparent, understandable and efficient classifiers. Most well known algorithms for their induction are based on a greedy divide-and-conquer strategy. This course will explore an alternative approach to their construction, namely evolutionary computation.

The evolutionary approach has benefits and drawbacks. The course will address both and suggest evasive actions for problematic points in evolutionary DT induction. Evolutionary principles can well be applied to other problems, too.

Lecturer

Dr. Matej Sprogar
University of Maribor, Slovenia
matej.sprogar@uni-mb.si

Dates

Date	Time	Room
We 16.10.2013	15:30 - 18:00	S2 219
Th 17.10.2013	15:30 - 18:00	KG 519
Fr 18.10.2013	13:45 - 16:15	S3 055

Introduction
Decision Tree taxonomy
Classification Trees
Regression Trees (Regression Trees, Model Trees)
Classic DT Tools
Evolutionary Computing
Genetic Algorithms
Genetic Programming
DT encoding
DT fitness
EC problems
DT evolution problems
Data preparation
Understanding Results
Summary

Exam

Students will have to do a project and send it to the lecturer. The marks for this course will be based on the project.

Topics for student projects:

EC crossover study: Implement the most typical EC crossover operators, together with theoretical and practical analysis of results
EC fitness functions study: Implement the most typical EC fitness functions, together with theoretical and practical analysis of results
EC mutations study: Implement the most typical EC mutation operators, together with theoretical and practical analysis of results
*EC engine for DT induction: Implement any EC for DT induction and apply it on a dataset of choice (UCI, ...)
**EC engine implementation +: Implement a generic EC engine in a language of choice, that is able to run either GA or GP and apply it on a chosen problem
**Self-adapting DT evolution: Implement a DT engine, which is able to self-adapt its parameters, and apply it on a chosen problem
**DT co-evolution: Implement a co-evolutionary DT engine, which evolves two competing populations, and apply it on a chosen problem
...

Literature