Prescriptive Analytics: Optimization and Simulation

Prescriptive Analytics: Optimization and Simulation Study Notes

Chapter Learning Objectives

• Extend analytics applications beyond descriptive (what happened) and predictive (what will happen) analytics to prescribe actions (what should we do).

• Implement techniques that can support complex decision-making processes, enabling organizations to optimize operations and achieve strategic goals.

• Techniques covered include mathematical programming, simulation, decision analysis, and genetic algorithms.

  • Use of spreadsheets or stand-alone software tools, often with specialized add-ins like Solver, for practical application of these techniques.

  • Familiarity with important concepts related to the underlying principles of prescriptive analytics and effective decision-making under various conditions.

  • Understand that modeling is a complex process requiring careful formulation and validation, closely related to the field of management science and operations research.

Key Understandings

• Recognize the limitations of predictive models when used in combination with prescriptive analytics techniques. Predictive models forecast outcomes, but they don't explicitly recommend actions; prescriptive models bridge this gap by suggesting optimal decisions based on those predictions.

8.1 Opening Vignette: Balancing Delivery Routes, Production Schedules, and Inventory

Problem Overview

• Company: Praxair/Linde, a global supplier of industrial gases, faces significant inventory management and logistics challenges due to the critical nature and high volume of their products.

• Logistics Director: Marcos Guimaraes oversees bulk distribution for White Martins in Brazil. His responsibilities include requiring efficient product allocation from plants, optimizing truck routing for deliveries, and managing inventory levels across a vast network for 150,000 tons of product/month across a fleet of 650 vehicles. This intricate operation ensures timely delivery of gases like oxygen and nitrogen to various industrial clients.

Challenges Identified

• Complex models with highly uncertain demand: Guimaraes faced the monumental task of coordinating over 18,000 delivery trips/month. The sheer volume and variability of demand necessitated a sophisticated approach to improve routing and achieve optimized distribution that traditional methods couldn't provide.

• Previous optimization tools lacked robust routing capabilities and struggled to integrate all necessary data points. The growing complexity of data sets, including real-time inventory, customer locations, and truck availability, led to significant inefficiencies and suboptimal decision-making.

Solution Approach

• Guimaraes opted for Analytic Solver, an advanced analytical software suite, specifically for its capabilities in managing logistics and supply chain optimization. This allowed for:

  • Daily forecasting of customer demand and three-month inventory balancing across multiple distribution centers and production plants.

  • Calibration for inherent uncertainty in demand fluctuation, incorporating reliability factors to minimize stockouts and overstocking.

• Outcome: The implementation led to an improved understanding of plant mechanics, production capacities, and operational costs. This comprehensive insight enabled more informed and better decision-making regarding product delivery schedules, resource allocation, and overall supply chain strategy, resulting in cost savings and improved service levels.

8.2 Model-Based Decision-Making

• Prescriptive Analytics: The utilization of advanced analytical models, often employing optimization, simulation, and heuristic techniques, to guide decision-making. These models leverage historical data, predictive forecasts, and constrained resources to recommend the best course of action.

• Initiating decisions based on customer predictions (e.g., promotional campaigns) and contractor selections (e.g., bidding strategies) can directly impact an organization's profitability, operational efficiency, and service quality.

Example Applications

• Awarding contracts to suppliers based on cost, reliability, and delivery time; calculating optimal advertising budgets to maximize reach and conversion; scheduling staff efficiently to meet demand while minimizing labor costs; optimizing delivery routes for logistics companies based on real-time traffic and demand forecasts; and determining optimal warehouse locations to minimize transportation costs and improve delivery times.

8.3 Structure of Mathematical Models for Decision Support

Mathematical models simplify real-world problems into quantifiable relationships, enabling systematic analysis.

Components of Decision Models

1. Decision Variables: These are the controllable inputs or choices that a decision-maker can directly influence. They represent the alternative courses of action available (e.g., $x<em>1, x</em>2$ represent the quantity of product A and B to produce; investment amounts in different assets).

2. Result Variables: These are the outputs or outcomes that reflect the effectiveness or performance of the system. They are typically dependent on the decision variables and uncontrollable variables (e.g., total profit, market share, customer satisfaction).

3. Uncontrollable Variables: Also known as parameters, these are factors or conditions external to the decision-maker's control but significantly affect the outcomes. They are typically given or estimated (e.g., interest rates, raw material costs, competitor actions, economic conditions, customer demand).

4. Intermediate Result Variables: These are variables that serve as a bridge, linking decision variables and uncontrollable variables to the final result variables. They often represent sub-calculations or intermediate stages within the model (e.g., total revenue calculated from unit price and quantity sold, which then contributes to the overall profit).

Mathematical Relationships

• Models intrinsically link these components through various mathematical expressions, which can be equations (establishing exact equalities) or inequalities (defining constraints or ranges). These relationships define how inputs transform into outputs based on the logic of the system being modeled.

Example Models

• Basic financial models, such as the profit equation: $P = R - C$, where P is profit, R is total revenue (e.g., Price x Quantity Sold), and C is total costs (e.g., Fixed Costs + Variable Costs x Quantity Produced). This equation shows a direct relationship between revenue, costs, and the ultimate profit.

8.4 Certainty, Uncertainty, and Risk

Decision-making environments can be classified based on the level of knowledge about future outcomes.

Types of Decision-Making Frameworks

• Decision-Making under Certainty: This framework assumes total knowledge of all relevant outcomes; conditions are deterministic, meaning that for each action, there is only one specific, known outcome. This leads to relatively easy model development and optimal solutions (e.g., investing in fixed-rate Treasury bills where the return is guaranteed).

• Decision-Making under Uncertainty: In this scenario, multiple possible outcomes exist for each decision, but the probabilities of these outcomes occurring are unknown or cannot be reliably estimated. This complicates model development as decision-makers rely on qualitative assessments, heuristics, or pessimistic/optimistic criteria (e.g., launching a new product in an unknown market with no historical data).

• Decision-Making under Risk: Similar to uncertainty, multiple outcomes are possible, but in this framework, known probabilities can be assigned to each outcome based on historical data, statistical analysis, or expert judgment. This allows for calculated risk assessments and the use of tools like expected value or decision trees (e.g., investing in mutual funds where historical performance provides probabilistic return ranges).

8.5 Decision Modeling with Spreadsheets

• Spreadsheets like Microsoft Excel, Google Sheets, or LibreOffice Calc enable powerful modeling through their extensive built-in statistical, financial, and logical capabilities. They are highly versatile and user-friendly.

• The importance of spreadsheets as end-user modeling tools continually increases due to their robust integration capabilities with databases (e.g., ODBC connections) and advanced analytical functions (e.g., array formulas, pivot tables).

• Add-ins like Solver (for optimization) or @RISK (for simulation) further extend their functionality, making them indispensable for many business analysts and managers for creating, analyzing, and presenting decision models.

8.6 Mathematical Programming Optimization

Key Concepts

• Mathematical Programming: A family of mathematical techniques used for the optimal allocation of scarce resources to optimize a measurable goal (e.g., profit maximization, cost minimization) subject to a set of constraints. Linear Programming (LP) is the most recognized and widely applied method when the objective function and all constraints are linear.

• Characteristics of allocation problems that make them suitable for mathematical programming include:

  • Limited quantity of resources, such as labor hours, raw materials, budget, machine time, or storage space.

  • A clear need for maximizing returns (e.g., profit, utility) or minimizing costs (e.g., production expenses, waste) on the allocated resources.

  • Constraints limiting allocation options, which can represent physical capacities, regulatory requirements, market demands, or budget limitations (e.g., a minimum production quota, a maximum available budget).

Example: Scheduling Physicians at the University of Tennessee Medical Center

• Problem: The medical center required an automatic and efficient method for scheduling physicians, especially in critical departments that operate 24/7. The challenge was to balance workload fairly among physicians while ensuring adequate coverage for all shifts, considering individual preferences and contractual agreements.

• Solution: A Hybrid Preference Scheduling Model (HPSM) was developed. This sophisticated model integrated various factors: physician availability, specific shift requirements (e.g., minimum and maximum number of doctors per shift), individual physician preferences (e.g., preferred days off, specific shift rotations), and workload equity. The HPSM likely employed integer programming or other optimization techniques to generate feasible and optimal schedules.

• Outcome: The implementation of the HPSM ensured better scheduling quality, significantly reducing manual effort and errors. It effectively accommodated individual preferences among physicians, leading to increased job satisfaction, while simultaneously guaranteeing comprehensive coverage for all critical shifts, thereby maintaining quality patient care and operational efficiency.

8.7 Multiple Goals, Sensitivity Analysis, What-If Analysis, and Goal Seeking

• Multiple Goals: In real-world management decisions, objectives are rarely singular. Various stakeholders (e.g., shareholders, employees, customers, regulators) often have competing goals that need to be considered simultaneously. Decision-makers must analyze these objectives, reconcile conflicts, and prioritize them, often requiring multi-objective optimization techniques.

• Sensitivity Analysis: This crucial technique examines how changes in input data (e.g., costs, demand, interest rates) affect the optimal solution or proposed outcomes. It is vital for understanding the robustness of a solution and adapting to dynamic environments and uncertain parameters.

  • Automatic Sensitivity Analysis: Provided by optimization software (like Solver), this feature quickly identifies the range within which input parameters can change without altering the optimal solution's structure or how variations in certain inputs (e.g., objective function coefficients, right-hand side of constraints) impact the optimal value of the objective function (e.g., shadow prices, reduced costs).

  • Trial-and-Error Sensitivity Analysis: Involves manually altering one or more input values in the model and observing the corresponding effects on the outputs. This iterative process helps in understanding the model's behavior and the impact of different assumptions.

Goal Seeking

• Goal-seeking techniques are reverse calculations that determine the necessary input values for achieving a desired output level. Instead of predicting an output given inputs, it finds the inputs needed for a specific output (e.g., finding the required sales volume to achieve a target profit margin; determining the necessary budget allocation to achieve a 15% growth in market share).

8.8 Decision Analysis with Decision Tables and Decision Trees

Decision Trees and Tables

• Decision Tables: These structured tools systematically list all possible decision alternatives, the states of nature (uncontrollable events), and the payoffs or consequences for each combination of decision and state of nature. They are particularly useful for making decisions under risk or uncertainty when the number of alternatives and states of nature is manageable.

• Decision Trees: These graphical tools represent decisions and their possible consequences in a sequential flow. They incorporate decision nodes (points where a choice is made), chance nodes (points where an uncertain event occurs with assigned probabilities), and terminal nodes (representing final outcomes or payoffs). Decision trees are excellent for complex, multi-stage decision problems involving sequential decisions and uncertainties, allowing for the calculation of expected values for each path.

8.9 Introduction to Simulation

Importance of Simulation in Decision Making

• Simulation allows for modeling complex real-world systems (e.g., factory operations, supply chains, financial markets) and conducting experiments on these models within a computer environment, without having to manipulate or disrupt the actual system. It is particularly effective for problems where analytical solutions are too complex or impossible due to a high degree of randomness, interdependence, and dynamic behavior. It handles uncertainty effectively by directly incorporating probabilistic elements.

• Types of simulation models:

  • Probabilistic Simulation: Also known as stochastic simulation, this type explicitly incorporates random variables using probability distributions (e.g., normal, uniform, exponential distribution) to represent uncertain events like demand fluctuations, service times, or arrival rates. It is typical in business scenarios such as inventory management, queuing systems, and project management where randomness is inherent.

  • Monte Carlo Simulation: A specific and widely used form of probabilistic simulation. It involves running multiple experiments (trials) of a model with randomly generated variables, drawn from their respective probability distributions. By repeating these trials thousands of times, it generates a distribution of possible outcomes, allowing decision-makers to understand the likelihood of various results, assess risk, and evaluate the impact of different strategies (e.g., simulating potential returns on an investment portfolio).

8.10 Genetic Algorithms and Developing GA Applications

Overview of Genetic Algorithms

• Genetic Algorithms (GA): A powerful class of optimization methods inspired by the principles of biological evolution and natural selection (survival of the fittest). GAs are particularly well-suited for solving complex optimization problems (e.g., non-linear, multi-modal, combinatorial) that are too intricate or computationally intensive for traditional analytical methods to solve efficiently or at all.

• Core Components that mimic biological processes:

  • Initialization: Creating an initial population of candidate solutions (chromosomes or individuals). This population is often generated randomly.

  • Fitness Function: A mechanism to evaluate the quality or