Performance Comparison of Search Algorithms for Manufacturing Systems using Discrete-event Simulation

1
Performance Comparison of Search Algorithms for
Manufacturing Systems using Discrete-event
Simulation

• Mohammed Yaseen Kalachikan Jafferali
• 09/13/02

2
Agenda

• Background
• Literature Review
• Motivation
• Problem Definition and Scope
• Experiments and Analysis
• Summary and Conclusion

3
1. Background

• Search Algorithms
• Simulation Optimization
• OptQuest Algorithm
• SimRunner Optimization Algorithm

4
1.1.Search Algorithms

• Helps to guide the model to a near-optimal
  solution.
• Local and Global Search Technique.
• Local techniques are myopic and get trapped in
  local optimum.
• Global Search techniques use special mechanisms
  to avoid local optima.

5
Contd
6
1.2. Simulation Optimization

• The process of finding the best input variable
  values from among all possibilities without
  explicitly evaluating each possibility (Carson et
  al, 1997).
• A Simulation Optimization Model

Feedback
Output
Input
Simulation Model
Optimization Algorithm
7
Contd

• Simulation Optimization Categories (Fu, 2001)

8
1.3. OptQuest Algorithm

• The software combines scatter search, tabu
  search, integer programming and neural networks
  into a single, composite search algorithm (Fu,
  2001).
• Scatter search is a population based approach
  like genetic algorithm which produces off springs
  more intelligently by incorporating search
  history of past evaluations (Glover et al, 1999).

9
Contd

• Tabu Search (TS) uses adaptive memory to prohibit
  the search from reinvestigating solutions that
  have already been evaluated and for guiding the
  search to a globally optimal solution. (Glover et
  al, 1999)
• Neural network is used to screen out solutions
  that are likely to be poor without allowing the
  simulation to evaluate them. (Glover et al, 1999)

10
1.4. SimRunner Optimization Algorithm

• The software combines evolutionary and genetic
  algorithms (Fu, 2001).
• Evolutionary algorithms operate on a population
  of potential solutions applying the principle of
  survival of the fittest to produce better and
  better approximations to a solution.

11
Structure of an Evolutionary Algorithm (GEA
Toolbox)
12
Contd
13
2. Literature Review

• Yunker et al, 1994 made a quantitative comparison
  between pattern search, response surface method
  and a genetic algorithm using simulation
  optimization on a university time-shared computer
  system.
• Lacksonen, 2001 compared pattern search, simplex,
  simulated annealing and genetic algorithm
  optimization algorithms on

14
Contd

• variations of four industrial case study
  problems.
• Fu et al, 1997 investigated the use of stochastic
  approximation for the optimization of a variety
  of discrete-event systems a single-server queue,
  a queueing network, and a transportation network
  via simulation.

15
3. Motivation

• Performance Comparison of popular search
  techniques have not been done explicitly for a
  more general configuration of a manufacturing
  system.
• Simulation optimization has been an active area
  of research.
• Genetic Algorithms, Tabu Search, Scatter Search
  and Neural Network are popular search techniques
  reported in literature.

16
Contd

• So commercial simulation optimization softwares
  (OptQuest for Arena and SimRunner for ProModel)
  implementing these techniques were used.
• Time constraints and difficulties involved in
  integrating simulation packages with search
  algorithms through computer codes.

17
4. Problem Definition and Scope

• This research compares the performance of
  OptQuest and SimRunner optimization algorithms to
  find the best set of queue dispatching rules, on
  different manufacturing system performances.

18
Contd

• The following performance measures were
  considered i) Proportion of Tardy Jobs ii)
  Total Throughput.
• The method used to assign due dates is the Total
  Work content method (TWK) Di Ri Ai, where,
  Ai k x Pi
• Queue disciplines considered FIFO, LIFO, SPT,
  LPT, EDD

19
Assumptions

• Model has a run time of 7200 min (8 hr production
  shift, 3 shifts everyday and 5 days a week).
• Material handling time between two machines is
  assumed to be negligible and the part routing
  time is assumed to be zero.
• Machine breakdown and maintenance activities are
  not considered.
• Batch size is assumed to be 1.

20
Assumptions (contd)

• In the stochastic case, the interarrival times
  and the processing times are sampled from an
  exponential distribution.
• In simulations that involve stochastic variables,
  the number of replications was assumed to be 5.
• The processing times of each job type is randomly
  generated from U10,50.

21
5. Experiments and Analysis

• Factors Affecting the Performance of a Search
  Algorithm
• Factors Considered in the Research
• Design of Experiments
• Analysis of Results

22
5.1. Factors Affecting the Performance of a
Search Algorithm

• System Complexity
• Nature of the variables (deterministic or
  stochastic)
• Computation Time
• Type of Variables (integer or real)

23
5.2. Factors Considered in the Research

• of part types (2, 10, 20)
• of machines (5, 12, 20)
• Capacity of queues (Finite, Infinite)
• Job Inter-arrival time (Deterministic,
  Stochastic)
• Processing time (Deterministic, Stochastic)
• Type of shop (Flowshop, Jobshop)

24
5.3. Design of Experiments

• Five factors, two of the factors at 3 levels, and
  the other factors at 2 levels each, constituting
  a total of 144 experimental runs to analyze the
  behavior of the system under full factorial
  design with 5 replicates each.

25
6. Conclusion

• Performance comparison of search algorithms was
  carried out for a more general manufacturing
  system rather than an adhoc case.
• Analysis of results will be carried out by
  running the experiments and by comparing the
  performance measures, the best search algorithm
  among the two will be reported.

26
Scope for Future Research

• More factors affecting the performance of the
  search algorithms like computation time, and the
  effect of different factors in manufacturing
  systems like material handling, batch size,
  machine breakdown etc. could be studied.
• More search algorithms could be compared.