Page 38 - Data Science Algorithms in a Week
P. 38

In: Artificial Intelligence                            ISBN: 978-1-53612-677-8
                       Editors: L. Rabelo, S. Bhide and E. Gutierrez   © 2018 Nova Science Publishers, Inc.






                       Chapter 2



                        USING DEEP LEARNING TO CONFIGURE PARALLEL

                            DISTRIBUTED DISCRETE-EVENT SIMULATORS



                                                                    2,
                                                    1
                                                                                       3
                                     Edwin Cortes , Luis Rabelo  and Gene Lee
                                                                      *
                                   1 Institute of Simulation and Training, Orlando, Florida, US
                                2 Department of Industrial Engineering and Management Systems,
                                       University of Central Florida, Orlando, Florida, US
                                3 Department of Industrial Engineering and Management Systems,
                                       University of Central Florida, Orlando, Florida, US


                                                        ABSTRACT


                              This  research  discusses  the  utilization  of  deep  learning  for  selecting  the  time
                          synchronization scheme that optimizes the performance of a particular parallel discrete
                          simulation hardware/software arrangement. The deep belief neural networks are able to
                          use  measures  of  software  complexity  and  architectural  features  to  recognize,  match
                          patterns and therefore to predict performance. Software complexities such as simulation
                          objects,  branching,  function  calls,  concurrency,  iterations,  mathematical  computations,
                          and messaging frequency were given a weight based on the cognitive weighted approach.
                          In  addition,  simulation  objects  and  hardware/network  features  such  as  the  distributed
                          pattern  of  simulation  objects,  CPUs  features  (e.g.,  multithreading/multicore),  and  the
                          degree  of  loosely  vs  tightly  coupled  of  the  utilized  computer  architecture  were  also
                          captured  to  define  the  parallel  distributed  simulation  arrangement.  Deep  belief  neural
                          networks  (in  particular  the  restricted  Boltzmann  Machines  (RBMs)  were  then  used  to
                          perform  deep  learning  from  the  complexity  parameters  and  their  corresponding  time
                          synchronization  scheme  value  as  measured  by  speedup  performance.  The  simulation



                       *  Corresponding Author Email: luis.rabelo@ucf.edu
   33   34   35   36   37   38   39   40   41   42   43