The Leibniz System Software is a versatile software package for logic
computation, construction of intelligent systems, data mining, machine learning,
and black box optimization. A key feature is that all models and results except
those of Black Box Optimization are expressed in logic formulas and thus are
mathematically compatible and also humanly comprehensible. This feature allows
seamless assembly of human knowledge and data mining or machine learning results
into humanly comprehensible models of complex processes.
The Leibniz System Software is comprised of several modules, each
developed by Dr. Klaus Truemper, Professor Emeritus of Computer Science. The Logic Computation and Construction
of Expert Systems module of the Leibniz System Software is available free of
charge from Dr. Truemper’s website at the University of Texas at Dallas for
purchasers of his book "Design of Logic-based Intelligent Systems". Since 2004, several additional modules
have been developed that are not part of the free distribution package. These
modules comprise several data mining and machine learning tools, plus an
optimization package for black box functions. Additional modules are planned.
UTD offers licenses of this expanded Leibniz System for a fee. The following modules are currently
available.
Logic Computation and Construction of Expert Systems – The module supports construction of intelligent systems that
seamlessly combine expert knowledge and logic relationships derived from prior
data, and that autonomously learn from feedback. It implements the theory and
algorithms of the books "Effective Logic Computation" (Wiley, 1998) and "Design
of Logic-based Intelligent Systems" (Wiley, 2004) by K. Truemper.
Machine Learning – The module learns logic relationships from data. The module not only
provides formulas for the separation of data sets, but also computes humanly
comprehensible explanations for differences in data. The module has been used in
a large number of application areas ranging from Economics, Engineering, and
Finance to Neurology and Medicine. Explanations produced by the module have been
used, for example, to devise investment strategies, criteria for credit
evaluation, and new treatments for cancer.
Subgroup Discovery – The module for Machine Learning described above requires that the
user has two or more distinct data sets and wants to know how and why these data
sets are different. In the Subgroup
Discovery case, there is just one data set, and the user wants to know whether
the data set has subsets of important characteristics. The Subgroup Discovery module detects
such subsets. A key aspect that differentiates the method from other subgroup
discovery methods is the use of so-called alternate random processes. This approach avoids the selection of
subsets whose structure may superficially seem important but actually is just a
random effect. The module has been employed in various settings. For example, a
blood serum marker for cervical cancer has been found that very reliably
predicts success/failure of current treatment methods. In a recent example, the module was used
to identify important subgroups of children with speech defects. In the cited cases, traditional
clustering methods failed to unearth these subgroups.
Optimization of Black Box Functions – There are an ever-growing number of applications where a model whose
objective function is not explicitly given, is to be optimized. For example, a simulation model may
involve several parameters that are to be optimized, and for the evaluation of
one set of values for these parameters the simulation model must be run. In effect, the simulation model is a
black box for computing values of the function to be minimized. If each simulation run takes some time,
say a few seconds up to hours, then any method for solving the optimization
problem must do so while evaluating just a few points for their objective
function value. The setting could
be rephrased by saying that the optimization problem is to be solved while using
a minimum number of function evaluations. The problem is even more difficult when
several objective functions must be simultaneously minimized. In that setting, the set of
pareto-optimal solutions is to be computed, again using a minimum number of
function evaluations. Here, a
solution is pareto-optimal if there is no other solution that is uniformly at
least as good as the given one, for all objective functions, and strictly better
for at least one of the objection functions.
The Optimization module solves such problems effectively, using a new
approach for which UTD has filed a patent application. So far, the method has been used in
several Engineering applications where pareto-optimal solutions were to be
computed. In each case, the method
found good solutions with very few function evaluations. For example, a problem of optimal
modeling in fluid dynamics was solved within one hour where competing methods
required at least one day.
Applications:
The software package is useful for applications involving or requiring,
simulation, computer modeling, or data analysis for underlying relationships.
Application areas include Economics, Engineering, Finance, Neurology, Medicine,
etc.
IP Status:
Patent Pending for method used in Black Box Optimization
module
Developer/Inventor:
Dr. Klaus Truemper, Professor Emeritus, Computer Science, The University
of Texas at Dallas
For Information Contact:
Becky Stoughton rebecca.stoughton@utdallas.edu
972-883-4574