TY - JOUR
T1 - Generating Optimal Decision Functions from Rule Specifications
AU - Gossen, Frederik
AU - Margaria, Tiziana
N1 - Publisher Copyright:
© 2017. All Rights Reserved.
PY - 2017
Y1 - 2017
N2 - In this paper we sketch an approach and a tool for rapid evaluation of large systems of weighted decision rules. The tool re-implements the patented mi-Aamics approach, originally devised as a fast technique for multicriterial decision support. The weighted rules are used to express performance critical decision functions. MiAamics optimizes the function and generates its efficient implementation fully automatically. Being declarative, the rules allow experts to define rich sets of complex functions without being familiar with any general purpose programming language. The approach also lends itself to optimize existing decision functions that can be expressed in the form of these rules. The proposed approach first transforms the system of rules into an intermediate representation of Algebraic Decision Diagrams. From this data structure, we generate code in a variety of commonly used target programming languages. We illustrate the principle and tools on a small, easily comprehensible example and present results from experiments with large systems of randomly generated rules. The proposed representation is significantly faster to evaluate and often of smaller size than the original representation. Possible miAamics applications to machine learning concern reducing ensembles of classifiers and allowing for a much faster evaluation of these classification functions. It can also naturally be applied to large scale recommender systems where performance is key.
AB - In this paper we sketch an approach and a tool for rapid evaluation of large systems of weighted decision rules. The tool re-implements the patented mi-Aamics approach, originally devised as a fast technique for multicriterial decision support. The weighted rules are used to express performance critical decision functions. MiAamics optimizes the function and generates its efficient implementation fully automatically. Being declarative, the rules allow experts to define rich sets of complex functions without being familiar with any general purpose programming language. The approach also lends itself to optimize existing decision functions that can be expressed in the form of these rules. The proposed approach first transforms the system of rules into an intermediate representation of Algebraic Decision Diagrams. From this data structure, we generate code in a variety of commonly used target programming languages. We illustrate the principle and tools on a small, easily comprehensible example and present results from experiments with large systems of randomly generated rules. The proposed representation is significantly faster to evaluate and often of smaller size than the original representation. Possible miAamics applications to machine learning concern reducing ensembles of classifiers and allowing for a much faster evaluation of these classification functions. It can also naturally be applied to large scale recommender systems where performance is key.
KW - Algebraic Decision Diagram
KW - Machine Learning
KW - MiAamics
KW - Optimal Decision Functions
KW - Rapid Decision Making
KW - System of Rules
UR - http://www.scopus.com/inward/record.url?scp=85127867476&partnerID=8YFLogxK
U2 - 10.14279/tuj.eceasst.74.1056.1030
DO - 10.14279/tuj.eceasst.74.1056.1030
M3 - Article
AN - SCOPUS:85050337925
SN - 1863-2122
VL - 74
SP - 1
EP - 16
JO - Electronic Communications of the EASST
JF - Electronic Communications of the EASST
ER -