Simulation And Modeling (ME-804) - Unit 2 Short Notes

UNIT 2: SIMULATION METHODOLOGY & TECHNIQUES

2.1 Input Data Analysis & Modeling

Purpose & Importance: Input data defines the stochastic elements of a model. Poor data leads to unreliable results (GIGO principle). The goal is to characterize real-world uncertainty with appropriate probability distributions.

Data Collection & Screening:

• Sources: Historical records, field observations, expert elicitation, synthetic data.

• Screening: Identify and handle outliers, data entry errors, and missing values. Use descriptive statistics (mean, std dev, min/max) and visual plots (box-and-whisker).

Distribution Identification:

1. Visual Methods:

   • Histogram: Shape (unimodal, skewed, symmetric).

   • Q-Q Plot (Quantile-Quantile): Compare data quantiles to theoretical distribution quantiles. Points following a straight line indicate a good fit.

   [!TIP] A Q-Q plot against a Normal distribution is a primary tool for checking normality.

2. Statistical Goodness-of-Fit (GOF) Tests:

   • Chi-Square ($$\displaystyle \chi^2 $$) Test: Bins data, compares observed vs. expected frequencies. Requires sufficient data per bin (≥5).

   • Kolmogorov-Smirnov (K-S) Test: Compares empirical cumulative distribution function (ECDF) to theoretical CDF. More powerful for continuous distributions, sensitive to median.

   • Anderson-Darling Test: A weighted K-S test, more sensitive to tails. Often preferred for Normality.

   • Hypothesis: $$\displaystyle H_0 $$: Data follows specified distribution. Reject $$\displaystyle H_0 $$ if p-value < $\alpha$ (e.g., 0.05).

Parameter Estimation:

• Method of Moments: Set sample moments equal to distribution moments.

  • Example (Normal): $$\displaystyle \hat{\mu} = \bar{x} $$, $$\displaystyle \hat{\sigma}^2 = s^2 $$.

  • Example (Exponential): $$\displaystyle \hat{\lambda} = 1 / \bar{x} $$.

• Maximum Likelihood Estimation (MLE): Finds parameters maximizing the probability of observed data. Often more efficient.

Multivariate & Time-Series Inputs:

• Correlation: Must be modeled if inputs are correlated (e.g., inter-arrival time and service time). Use copulas or multivariate distributions.

• Time-Series: For sequential dependent data (e.g., daily demand). Models include ARIMA (AutoRegressive Integrated Moving Average).

2.2 Random Number & Variate Generation

Foundations of Randomness: A good Pseudo-Random Number (PRN) sequence must be:

1. Uniformly distributed over (0,1).

2. Independent (no serial correlation).

3. Have a long period (before sequence repeats).

4. Be repeatable (reproducible with same seed).

Pseudo-Random Number Generators (PRNGs):

• Linear Congruential Generator (LCG):

$$X_{n+1} = (aX_n + c) \mod m$$

*   $$\displaystyle X_0 $$ = seed, $a$ = multiplier, $c$ = increment, $m$ = modulus.

*   $$\displaystyle U_n = X_n / m $$ gives uniform (0,1).

*   **Period** depends on choice of $a, c, m$. Full-period possible if: $c$ and $m$ relatively prime, $a-1$ divisible by all prime factors of $m$, $a-1$ divisible by 4 if $m$ divisible by 4.

*   **Issue:** Can exhibit serial correlation on a lattice.

• Combined Generators (e.g., L'Ecuyer): Combine two or more LCGs to achieve a longer period and better statistical properties.

Testing PRNGs (Empirical Tests):

• Frequency Test: Check counts in intervals for uniformity ($$\displaystyle \chi^2 $$ test).

• Runs Test: Check sequences of above/below mean for randomness.

• Autocorrelation Test: Check correlation between $$\displaystyle U_n $$ and $$\displaystyle U_{n+k} $$.

• Gap Test: Check distances between occurrences in a specified range.

• Poker Test: Check frequencies of digit patterns (e.g., "pairs" in 5-digit numbers).

Generating Random Variates:

1. Inverse Transform Technique:

   • Method: If $U \sim \text{Uniform}(0,1)$ and $F$ is CDF of target distribution, then $$\displaystyle X = F^{-1}(U) $$ has the target distribution.

   • Applications:

     • Exponential($\lambda$): $$\displaystyle X = -\frac{1}{\lambda} \ln(1-U) $$

     • Uniform($a,b$): $$\displaystyle X = a + (b-a)U $$

     • Weibull($\alpha,\beta$): $$\displaystyle X = \frac{1}{\beta} (-\ln(1-U))^{1/\alpha} $$

   • Limitation: Requires invertible CDF.

2. Acceptance-Rejection (A-R) Technique:

   • Method: Generate $Y$ from proposal $g(x)$ and $U$. Accept $Y$ if $U \leq f(Y)/(c g(Y))$, where $f$ is target PDF, $$\displaystyle c \geq \sup_x f(x)/g(x) $$.

   • Efficiency: Probability of acceptance = $1/c$. Goal is small $c$.

   • Applications: Normal (using Exponential proposal), Gamma, Beta.

3. Convolution: Sum of independent random variables. Example: Erlang($k,\lambda$) = sum of $k$ i.i.d. Exponential($\lambda$) variables.

4. Special-Purpose Generators:

   • Box-Muller Transform (Normal):

$$Z_1 = \sqrt{-2\ln U_1} \cos(2\pi U_2), \quad Z_2 = \sqrt{-2\ln U_1} \sin(2\pi U_2)$$

    where $$\displaystyle U_1, U_2 \sim \text{Uniform}(0,1) $$. $$\displaystyle Z_1, Z_2 \sim N(0,1) $$.

2.3 Model Verification & Validation (V&V)

[!TIP] Common Pitfall: Confusing verification (correct code) with validation (correct model). Always verify first.

2.4 Experimentation & Output Analysis

Statistical Nature of Output: Simulation output is a random variable. Must use statistical methods for analysis.

Types of Simulations:

• Terminating: Natural end point (e.g., one day, one project). Output is a finite set of observations.

• Non-Terminating (Steady-State): No natural end; interest is in long-run average. Requires handling initial transient (warm-up period).

Output Analysis for Terminating Simulations:

• Single Replication: Use all observations from one run. Point estimator $\bar{Y}$, confidence interval:

$$\bar{Y} \pm t_{\alpha/2, n-1} \frac{S}{\sqrt{n}}$$

where $S$ = sample std dev, $n$ = sample size.

• Multiple Replications: Run $k$ independent replications. Use batch means (treat each replication's mean as one observation) to form CI.

Output Analysis for Steady-State Simulations:

1. Warm-up Period Determination: Must delete initial transient to avoid bias.

   • Welch's Method: Plot moving averages of means from multiple replications; look for stabilization.

   • Graphical: Plot time series of output variable; identify stabilization point.

2. Analysis Methods:

   • Replication-Deletion: Run $k$ replications, delete warm-up from each, use replication means as independent observations.

   • Batch Means (Single Long Run): Divide post-warm-up data into $m$ large batches. Treat batch means as approximately independent. CI:

$$\bar{Y} \pm t_{\alpha/2, m-1} \frac{S_b}{\sqrt{m}}$$

    where $$\displaystyle S_b $$ = std dev of batch means.

*   **Key Requirement:** Batches must be large enough for batch means to be nearly independent (often $m \geq 10$).

Comparing System Alternatives (Scenarios):

• Paired-t Method (with Common Random Numbers - CRN):

  • Use same random number streams (seeds) for each alternative to induce positive correlation.

  • For $k$ replications, compute differences $$\displaystyle D_i = Y_{i,A} - Y_{i,B} $$.

  • CI for difference $$\displaystyle \mu_D $$: $$\displaystyle \bar{D} \pm t_{\alpha/2, k-1} \frac{S_D}{\sqrt{k}} $$.

  • Advantage: Reduces variance, fewer replications needed.

• Independent Replications Method: Run separate replications for each alternative. Use two-sample t-test (assuming equal/unequal variances).

Design of Simulation Experiments:

• OFAT (One-Factor-At-a-Time): Change one input factor at a time. Inefficient; misses interactions.

• Factorial Designs: Systematically vary multiple factors. Full factorial tests all combinations. Fractional factorial tests a subset.

• Response Surface Methodology (RSM): Fit a polynomial (e.g., quadratic) model to response vs. factors to find optimal settings.

2.5 Introduction to Simulation Software & Experiment Management

Role of Simulation Software:

• General-Purpose Simulators: AnyLogic, Simio, Arena. Provide library of modeling constructs (queues, resources, flows), input processors, experiment managers, animators, output analyzers. Faster development, built-in statistics.

• General-Purpose Programming: Python (SimPy), R, Java, C++. More flexibility, control, and scalability for complex logic or custom algorithms.

Designing Experiments in Software:

• Key Parameters to Set:

  • Replication Length: How long each run lasts.

  • Number of Replications: For statistical reliability.

  • Warm-up Period: For steady-state models.

  • Output Collection: Specify what variables/statistics to track (e.g., queue length, resource utilization, time-in-system).

Documentation & Reporting:

• Model Documentation: Assumptions, input distributions & parameters, verification/validation steps, experiment design.

• Results Reporting: Clear presentation of scenarios compared, statistical methods used (CI, test type), confidence intervals, and practical significance of differences. Include run specifications (seed, length, replications) for reproducibility.

[!TIP] Always report both statistical and practical significance. A difference may be statistically significant but not operationally meaningful.