Download as ppt, pdf, or txt
Download as ppt, pdf, or txt
You are on page 1of 15

c 


  
  

‡ Instructor: Giampiero Pecelli


‡ e-mail: giam@cs.uml.edu
‡ Office Phone: 978 - 934 -3639
‡ Office: Olsen 225
‡ Office Hours: Before Class and by Appointment
c  
  
  

‡ º  
    
‡  
    
‡ º   
    
c  
  
  

‡ WHY?
We need to conduct experiments "on some reality" and the
reality - although pre-existing - is not available for our
experiments.
Examples:
a) a busy network of computers that cannot be taken over just
for the experiment;
b) a busy superhighway system on which we want to "change
the rules of traffic";
c) a chemical plant whose production cannot be stopped so that
"we can tinker with it"; etc..
c  
  
  
‡What characteristic do these example share?
They simply have to do with our lack of  to an existing
artifact: the simulation allows us to construct a useful model of
the artifact, that we can then use as though it were the
inaccessible artifact. The goal is to determine whether a
planned change to the USE of the artifact can be implemented
while producing the desired results and no undesired ones.

A more specific example would be the introduction of the use of


a "group productivity package", like Lotus Notes or a
Configuration and Version Manager for a software producing
organization. In both cases the traffic patterns - and
bottlenecks - in a LAN might not be predictable without
extensive testing, and any meaningful REAL testing will result in
many lost productivity hours for the whole group or organization
that is adopting the package.
c  
  
  

‡ A second set of examples.


These have to do with the   of an appropriate artifact.
Here are some examples:
a) An automobile frame that must meet certain stiffness and
crushability criteria, while also meeting geometry, materials,
production method and weight constraints;
b) An algorithm to manage certain types of (not yet available?)
traffic in networks with as yet non-existent (but likely, or already
possible) properties (e.g., 20 TH bandwidth);
c) The design of drugs with special properties;
c  
  
  

‡ What characteristic do these example share?


There is NO artifact on which to perform experiments, and the
construction of any such artifact is not feasible (too expensive -
current technology is too immature - too dangerous) without
knowledge that the finished artifact will behave (with high
probability) as desired.
There MAY exist earlier versions of similar artifacts, with different
characteristics, that MIGHT be used as guides for the design of a
simulation, but with no guarantee that the results of the
simulation can be compared to "real" data in the regions of
interest.
c  
  
  

‡   (Shannon):

The process of designing a computerized model of a   (or


) and conducting experiments with this model for the
purpose either of understanding the behavior of the system or of
evaluating various strategies for the operation of the system.

 : an orderly collection of logically related principles, facts


or objects.
: a method of doing something involving multiple steps
and operations.
c  
  
  

‡   


A)  
  : the collection of external factors
capable of causing a change in the system.
B)    : the minimal collection of information with
which the future behavior of a system can be reliably (uniquely?)
predicted.
C)    : any events that causes a State Change.
D)
    : one occurring   the system.
E)
    : one occurring   the system.
c  
  
  

´)     : one in which the quantities of interest


are represented by continuous variables (e.g., distances between
cars on a highway).
G)    : one in which the quantities of interest are
represented by integer-valued variables (e.g., number of cars on
a highway).
´)   : one in which both integer and continuos
variables appear (e. g., number of and distances between cars)
and are of interest.
c  
  
  

G)      : one in which the next state is


    by the current state. Examples: Classical
Mechanics; anything that can be adequately modeled via
Newtonian Mechanics: hit the brakes of your car under exactly
controlled conditions and the distance it takes for you to come to
a stop can be exactly predicted. Deterministic Automata.
H)     : one in which the next state is only
    by the current state - there are
multiple possible next states that can occur subsequent to the
same activity, each with a given probability. Examples: Quantum
Mechanical phenomena. Non-deterministic automata.
c  
  
  

        .


‡   
a) !  : what is it and what do I want
to do with it?
b) "
   : time, people and money.
c)     
‡# 
a) #$ : find relationships.
b)    : find and collect appropriate data.
c) #    : program and debug.
c  
  
  
‡  % 
a)   : does the PROGRAM   as
intended?
b)   : does the PROGRAM    as
intended?
‡ 
a)
   : run it!
b)  : how do I analyze and interpret the results?
c) &   %   : how do I
implement the decisions resulting from the simulation,
and how do I document the model and its use?
c  
  
  
‡  #.
What is it that we are measuring?
What (statistical) properties of the "measured" are we
interested in?
´or example: maximum, minimum, totals, mean, variance, higher
moments, specific frequency distribution, interarrival times,
service times, lengths of queues, loss rates, error rates, etc.
c  
  
  
Advantages of Simulation.
a) it permits controlled experimentation.
you KNOW what parameters are being changed.
b) it permits time compression.
e.g., weather forecasting...
c) it permits sensitivity analysis (change input vars)
d) it does not disturb the real system.
which may not even exist, anyway.
e) it is an effective training tool.
you are not likely to crash a flight simulator,
or a big chunk of the Internet.
c  
  
  
My own interest:
How do you experiment in a meaningful way with
algorithms whose theoretical properties can be predicted?
Most of these algorithms attempt to provide management
for traffic which is, as yet, not well understood, in networks with
characteristics that don't yet exist.
Qualitative and quantitative mathematical predictions can
be obtained only under considerably simplified assumptions on
the system being studied. How well will these predictions
compare with reality? Can simulation provide a reasonable
answer?
Many engineers construct an algorithm that will exhibit
SOME desired behaviors, run a few simulations, call it quits and
send a paper out. Is this a prescription for nonsense? Or worse?

You might also like