SIMULATION LAB (EX-804) - Unit 5 Short Notes

UNIT 5: ADVANCED SIMULATION CONCEPTS & APPLICATIONS

5.1 Advanced Modeling Paradigms

5.1.1 Discrete-Event Simulation (DES) Deep Dive

• Core Idea: Models systems as a sequence of discrete events over time. State changes only at event times.

• Advanced Entity Routing: Uses decision nodes, routing expressions, and conditional logic (e.g., IF-THEN-ELSE, SELECT) to direct entities based on attributes or system state.

• Resource Management: Models resource failures (breakdowns) and preventive maintenance using schedules or failure distributions (e.g., Weibull). Resource sets and capacity pools handle complex sharing.

• Scheduling Logic: Implements priority-based dispatching rules (e.g., FCFS, SPT, LPT, EDD) and sequencing for queues and resources.

[!TIP] Exam Focus: Be prepared to model a system with a machine that has a random time-between-failures (e.g., Weibull distribution) and a scheduled maintenance every 100 hours.

5.1.2 Agent-Based Modeling (ABM)

• Core Idea: Models a system as a population of autonomous agents with individual behaviors and interactions.

• Agent Components:

  • Attributes: Static/dynamic properties (e.g., age, health, wealth).

  • Decision Rules: IF-condition logic based on attributes, environment, or other agents.

  • Interactions: Communication, collision, competition, or cooperation protocols.

• Emergent Phenomena: Global system patterns (e.g., traffic jams, market trends) that arise from simple local agent rules. Key output is population-level statistics.

5.1.3 System Dynamics (SD)

• Core Idea: Models a system using continuous flows between stocks (accumulations) governed by feedback loops.

• Key Elements:

  • Stock: Level of something (e.g., Inventory, Population).

  • Flow: Rate of change into/out of a stock (e.g., Arrival Rate, Birth Rate).

  • Converter: Parameters or auxiliary variables (e.g., Growth Rate).

  • Connector: Shows influence between elements.

• Causal Loop Diagrams (CLD): Visual tool to identify reinforcing (R) and balancing (B) feedback loops.

  • R-loop: Amplifies change (e.g., Viral spread).

  • B-loop: Stabilizes system (e.g., Inventory reordering).

5.1.4 Hybrid/Multi-Method Simulation

• Concept: Integrates two or more paradigms (e.g., DES + ABM, SD + DES) in one model to leverage their strengths.

• Use Case Example: A supply chain where:

  • SD models national inventory levels (continuous flows).

  • DES models individual warehouses and truck dispatch (discrete events).

  • ABM models consumer buying behavior and retailer decisions (autonomous agents).

• Challenge: Requires sophisticated software (e.g., AnyLogic) and careful interface design between sub-models.

5.2 Input Modeling & Data Analysis

5.2.1 Data Collection and Fitting

1. Identify Distribution: Use histograms and Q-Q plots to visually assess data shape (e.g., Exponential for inter-arrivals, Normal for task times, Lognormal for skewed positive data, Weibull for time-to-failure).

2. Parameter Estimation: Calculate parameters (e.g., $\mu, \sigma$ for Normal; $\beta, \theta$ for Weibull) from sample data.

3. Goodness-of-Fit Tests:

   • Chi-Square ($$\displaystyle \chi^2 $$) Test: Bins data, compares observed vs. expected frequencies. Sensitive to bin choice.

   • Kolmogorov-Smirnov (K-S) Test: Compares empirical CDF to theoretical CDF. More powerful, no binning.

   • Anderson-Darling (A-D) Test: Similar to K-S but gives more weight to tails. Often preferred.

4. Interpretation: Use p-value. If $$\displaystyle p > \alpha $$ (e.g., 0.05), fail to reject the hypothesis that data fits the distribution. Do not accept it as true—lack of evidence against ≠ proof for.

5.2.2 Input Uncertainty and Sensitivity

• Impact: Poor distribution choice can lead to misleading output (e.g., underestimating queue lengths).

• Sensitivity Analysis: Systematically vary key input parameters (e.g., arrival rate mean, service time std dev) within plausible ranges to see effect on Key Performance Indicators (KPIs).

• Scenarios: Define "Best Case," "Worst Case," and "Most Likely" input sets to bound output variability and assess risk.

[!TIP] Common Pitfall: Using a distribution just because it's "common" (e.g., Normal for all data). Always fit and test.

5.3 Output Analysis & Statistical Techniques

5.3.1 Types of Output Data

• Transient Behavior: System output during the initial "start-up" phase. Not representative of long-run operation.

• Steady-State Behavior: Long-run, stable performance of a non-terminating system.

• Terminating Simulation: Has a natural end (e.g., a 1-year project). All output is relevant.

• Non-Terminating Simulation: Runs indefinitely (e.g., a 24/7 call center). Must discard transient period (warm-up).

5.3.2 Statistical Validity

• Warm-Up Period: Initial time period to discard to achieve steady-state.

  • Welch's Method: Run multiple replications with different random number seeds. Plot moving average of output (e.g., WIP). The point where curves stabilize is the warm-up period.

• Autocorrelation: Successive observations in a time series are not independent (common in DES). Violates assumption of standard statistical tests.

  • Solution: Use batch means (divide output into large batches, treat batch means as independent) or replication-deletion (run many independent replications, discard warm-up from each).

5.3.3 Comparing System Configurations

• Design: For comparing two systems (A vs. B), use paired-t test on replication means (same random numbers for both systems).

  • Confidence Interval (CI) for difference ($$\displaystyle \mu_A - \mu_B $$):

$$ \left( \bar{d} - t_{\alpha/2, n-1} \frac{s_d}{\sqrt{n}}, \bar{d} + t_{\alpha/2, n-1} \frac{s_d}{\sqrt{n}} \right) $$

where $\bar{d}$ = mean of paired differences, $$\displaystyle s_d $$ = std dev of differences, $n$ = number of replications.

*   If CI **does not contain 0**, difference is statistically significant.

• For >2 Systems: Use Analysis of Variance (ANOVA) on replication means.

5.4 Model Verification & Validation (V&V)

[!TIP] V&V is iterative. Validation failures often lead to model changes, requiring re-verification.

5.5 Advanced Simulation Software Features

• Custom Programming: Extend built-in logic using VBA (Arena), Simio Processes (C#/VB), or Java (AnyLogic). Create user-defined functions for complex calculations.

• I/O Data Management:

  • Read: Import external data (CSV, Excel, Databases) for arrival schedules, resource schedules, or time-varying arrivals.

  • Write: Export detailed transaction logs or time-series output for external analysis (e.g., in R, Python).

• Animation & 3D Visualization:

  • Debugging: Watch entity flow to spot logic errors.

  • Communication: Create realistic 3D layouts (factories, hospitals) to convey model and results to non-technical stakeholders.

• Experiment Manager & Optimization:

  • Design of Experiments (DOE): Systematically run model under thousands of input combinations (factorial, Latin Hypercube).

  • Optimization: Link to solvers (e.g., OptQuest in Simio, SimRunner in Arena) to automatically search for input settings that maximize/minimize a KPI (e.g., minimize cost, maximize throughput).

5.6 Simulation in the Analysis Lifecycle

1. Problem Formulation:

   • Define clear objective (e.g., "Reduce average patient wait time by 20%").

   • Set scope (what's in/out) and KPIs (e.g., Avg. Wait Time, Resource Utilization, Throughput).

   • Develop conceptual model (block diagram).

2. Experimentation & Interpretation:

   • Design experiments (e.g., 2^k factorial design to study k factors).

   • Analyze main effects and interactions.

   • Translate stats into business insights: "Adding one nurse reduces wait time by 15 minutes, but only if physician schedule is also adjusted."

   • Perform risk analysis: "There is a 30% probability that the new design will fail to meet the 30-minute target."

3. Implementation & Follow-up:

   • Create executive summary with clear recommendations, visualizations (animation, charts), and discussion of model limitations.

   • Plan for model reuse (document assumptions, code, data sources).

   • Recommend pilot study or A/B test to validate model predictions in real world.

5.7 Case Studies & Application Domains

5.8 Future Trends & Ethical Considerations

5.8.1 Integration with Other Technologies

• Digital Twin: A live, bidirectional virtual replica of a physical system. Simulation is the core engine. Real-time data feeds update the model; model predictions guide physical system control.

• AI/ML Integration:

  • ML for Input Fitting: Automatically select best-fit distributions.

  • AI Agents in ABM: Use reinforcement learning to train agents to make optimal decisions.

  • Surrogate Modeling: Use ML to create fast approximations of slow simulation models for optimization.

• Cloud-Based Platforms: Enable collaborative modeling, massive parallel runs, and scalable computing (e.g., Simio Cloud, AnyLogic Cloud).

5.8.2 Model Credibility and Ethics

• Communicating Uncertainty: Always present results with confidence intervals or probability distributions, not just single point estimates. State assumptions and limitations clearly.

• Avoiding Bias:

  • Data Bias: Ensure input data is representative (no sampling bias).

  • Assumption Bias: Question if model logic inadvertently favors a stakeholder's agenda.

  • Algorithmic Bias: In ABM with AI agents, check if learned behaviors are fair/equitable.

• Ethical Use:

  • Do not use simulation to justify harmful decisions (e.g., cutting safety staff) without rigorous validation.

  • Be transparent about model's purpose and scope to prevent misinterpretation.

  • Consider broader societal impacts (e.g., simulating traffic flow might displace communities).

[!TIP] In exams, always mention validation and uncertainty communication when discussing the final implementation stage. It shows depth.