A
Symbolic Logic Approach of Deriving Initial Neural Network Configurations for
Supervised Classification
Shie-Jue
Lee, Mu-Tune Jone
Abstract.
One of the problems encountered in neural network applications is the choice of
a suitable initial neural network configuration for the given classification
problem. We propose an idea of constructing initial neural network
configurations by making use of decision trees and threshold logic. First, a
decision tree is constructed from the given set of training patterns. Then the
decision tree is translated into a neural network. Initial values for the
weights and thresholds of the neural network are determined. Finally, the
obtained neural network is trained by the back-propagation algorithm.
Experimental results have shown that a neural network constructed in this manner
learns fast and performs efficiently.
Algorithm
Mapping with Parallel Simulated Annealing
B.
Robič, J. Šilc
Abstract.
This paper presents a parallel simulated annealing algorithm for solving the
problem of mapping irregular parallel programs onto homogeneous processor arrays
with regular topology. The algorithm constructs and uses joint transformations.
These transformations guarantee
a high degree of parallelism that is bounded below by
where |Np| is the number of
task nodes in the mapped program graph Gp and deg(Gp)
is the maximal degree of a node in Gp.
The mapping algorithm provides a good program mappings (in terms of program
execution time and the number of processors used) in a reasonable number of
steps.
Solving
the General Linear Model on a SIMD Array Processor
E.J.
Kontoghiorghes, M.R.B. Clarke
Abstract.
Two parallel algorithms are proposed for the solution of the General Linear
Model on a SIMD array processor. The first algorithm employs efficiently
compound Givens rotations while the second algorithm uses Householder
transformations. The implementation of the two algorithms on the 1024 processor
AMT DAP-510 is described and their performance analysed using high accurate
execution time models. No single algorithm is superior in all ranges examined
and the best choice depends on the problem size and the number of processing
elements available.
Performance
Prediction and Expert Adviser for Automatic Parallelisation of Fortran Programs
R.
Blaško
Abstract.
The peak processing performance of highly parallel computers can be achieved
only by advanced programming environments and tools developed particularly for
such systems. Our global aim is to develop tools for an automatic
parallelization of the Fortran programs, in a framework of Vienna Fortran
Compilation System. The proposed parallelization system consists of the three
kernel subsystems: transformation system, performance prediction system, and
expert adviser. This paper is devoted to the performance prediction system and
expert adviser designed as the new tools promoting automatic parallelization.
The performance prediction system derives performance characteristics
from the sequential and parallel versions of the program, during parallelization
by the transformation system, i.e. before execution on the parallel computer.
Performance prediction results are utilized by a user, directly, or by the
expert adviser. The expert adviser guides the user through a program improvement
process. Both new subsystems are integrated with the transformation system,
creating an advanced parallel programming environment.
The
Distributed Multimedia Learning Environment Employing Gaming-Simulation Method
with Expert Systems in the World of Macro Economics
T.
Okamoto, Y. Ueda, M. Kunishige
Abstract.
This study aims at describing the architecture of the distributed multimedia
learning environment employing gaming/simulation method for the world of
"Macro Economics". Especially, this system is emphasized on the
integrated framework of the advanced technologies like the higher
gaming/simulation, digitized image processing, expert system, communication
technology of file serving, in consideration of the concept of intelligent
learning environment. In the system, each of the participating students can play
the role of an agent in "macro economics world" on the basis of social
situated learning by using the function of gaming/simulation under the
networking environment. The system incorporates an expert system representing
the virtual smart student and a chairman-expert system which controls the whole
system. The idea is derived from the concept according to which the computer
companion should be embedded in the system in order to support the student in
learning by observation and modeling.