Prescriptive Analytics Systems Flashcards

Question 1

What does prescriptive analytics aim to determine for a decision maker, and how does it differ from prediction?

Accepted Answer

Prescriptive analytics aims to recommend the best action or set of actions for a given decision context. Predictive analytics estimates what is likely to happen, while prescriptive analytics builds on those estimates to evaluate alternatives against an objective and constraints, then suggests which choice to make.

Question 2

How do descriptive, predictive, and prescriptive analytics differ in the questions they answer?

Accepted Answer

Descriptive analytics asks what happened, predictive analytics asks what is likely to happen, and prescriptive analytics asks what should be done about it. Each stage typically uses the output of the previous one, so a prescriptive system depends on reliable descriptions and forecasts. This sequence helps you choose the right tool for a question.

Question 3

What is a decision variable, and why is it central to a prescriptive model?

Accepted Answer

A decision variable is a quantity that the decision maker can control, such as how many units to produce or which route to assign to a vehicle. The solver chooses values for these variables, so every recommendation is ultimately a set of values assigned to them. Clear variable definitions prevent later confusion.

Question 4

What is an objective function in a prescriptive model?

Accepted Answer

An objective function is a mathematical expression that measures how good a candidate decision is, such as total profit, total cost, or delivery time. The model then maximizes or minimizes it. Because the objective encodes what the decision maker values, a poorly chosen objective can produce a technically optimal answer that is practically useless.

Question 5

What are constraints, and how do they limit the set of decisions a model may consider?

Accepted Answer

Constraints are conditions that a feasible decision must satisfy, such as capacity limits, budget caps, or required service levels. Together they define the feasible region, which is the set of decisions that respect every rule. The optimal solution is then the best point inside that region, so editing one constraint can change the recommendation.

Question 6

What is a feasible solution, and how does it differ from an optimal solution?

Accepted Answer

A feasible solution is any set of decision values that satisfies all constraints, even if it performs poorly. An optimal solution is a feasible solution that gives the best objective value among all feasible alternatives. Many feasible solutions may exist, but only the best one is recommended, so both properties must hold.

Question 7

Why do prescriptive models often need a sensitivity analysis after the solution is found?

Accepted Answer

Sensitivity analysis tests how much the recommendation changes when inputs such as costs, demand, or capacity are varied. Real data are estimates, so a solution that is fragile to small input changes may be risky to implement. Sensitivity results also show which constraints are binding and which inputs most affect the outcome.

Question 8

What is a binding constraint, and what does it signal about the optimal solution?

Accepted Answer

A binding constraint is one that holds with equality at the optimal solution, meaning the limit is fully used. It signals that relaxing that limit could improve the objective, which is why extra capacity has value. A nonbinding constraint has slack, so small changes to its limit leave the recommendation unchanged.

Question 9

What is linear programming, and what two features make a problem linear?

Accepted Answer

Linear programming is a method for optimizing a linear objective function subject to linear constraints over continuous decision variables. A problem is linear when the objective and every constraint are sums of variables multiplied by fixed coefficients, with no products or powers of variables. This structure lets solvers find global optima efficiently.

Question 10

Why are integer variables used in optimization models, and what makes them harder to solve?

Accepted Answer

Integer variables are used when decisions must be whole numbers, such as the number of machines to buy. They model yes or no choices through binary variables that take only the values zero or one. Integer restrictions make problems harder because the solver must search among discrete options.

Question 11

What is the difference between integer programming and mixed integer programming?

Accepted Answer

Integer programming requires every decision variable to take an integer value, while mixed integer programming allows some variables to be integers and others to be continuous. Many practical planning models are mixed, combining continuous quantities with yes or no choices. The mixed form is more flexible than pure integer programming.

Question 12

How does the big M method represent a logical either or condition in a linear model?

Accepted Answer

The big M method adds a binary variable and a very large constant so that a constraint can be switched on or off. When the binary variable is zero, the large constant makes the constraint automatically satisfied, and when it is one, the original constraint must hold.

Question 13

What is a network flow model, and which real problems is it commonly used to represent?

Accepted Answer

A network flow model represents a system as nodes connected by arcs, where each arc carries a quantity of flow subject to capacity limits. Shipping goods from warehouses to customers, routing data through communication links, and moving materials through supply chains are common examples. Their special structure often allows fast solutions.

Question 14

What is a shortest path problem, and how is it formulated as a decision model?

Accepted Answer

A shortest path problem seeks the route between two nodes that minimizes total cost, distance, or time. In a decision model, each arc gets a binary variable for use, flow balance constraints send one unit from origin to destination, and the objective sums arc costs.

Question 15

What is a traveling salesperson problem, and why is it considered computationally hard?

Accepted Answer

The traveling salesperson problem asks for the shortest tour that visits each city exactly once and returns to the starting point. It is considered computationally hard because the number of possible tours grows rapidly as cities are added, so brute force fails quickly. Exact methods work for moderate sizes, while heuristics are common for larger instances.

Question 16

How does a mathematical model differ from a spreadsheet calculation in prescriptive work?

Accepted Answer

A mathematical model states decision variables, an objective, and constraints explicitly, so a solver can search for the best feasible values. A spreadsheet calculation evaluates one scenario that a person has chosen, and the user must try alternatives by hand. Spreadsheets remain useful for prototyping, but optimization models scale better to large choice spaces.

Question 17

What is stochastic optimization, and how does it handle uncertain inputs?

Accepted Answer

Stochastic optimization is a family of methods that choose decisions while explicitly modeling random inputs, such as demand or prices, through probability distributions or scenarios. The model typically evaluates each candidate decision by its expected performance across the scenarios. This lets the recommendation hedge against variability instead of assuming a single forecast is correct.

Question 18

What is a scenario analysis, and when is it preferred over full stochastic optimization?

Accepted Answer

Scenario analysis evaluates a decision under a small set of plausible future states, such as optimistic, expected, and pessimistic market conditions. It is preferred when the number of uncertain factors is limited, when probabilities are hard to estimate, or when decision makers mainly want to see how outcomes change across contrasting futures.

Question 19

What is Monte Carlo simulation, and what steps does it follow?

Accepted Answer

Monte Carlo simulation estimates the behavior of a system by repeatedly sampling random inputs and recording the resulting outcomes. The usual steps are to define the model logic, specify probability distributions for uncertain inputs, generate many random samples, compute the output for each sample, and summarize the results with averages, percentiles, or probability estimates.

Question 20

How is a probability distribution used to represent uncertainty in a model input?

Accepted Answer

A probability distribution assigns likelihoods to the possible values of an uncertain input, such as the daily demand for a product. Analysts choose a distribution based on historical data, expert judgment, or both, and then sample from it during simulation. Choosing a distribution that poorly fits the data can lead to misleading recommendations.

Question 21

What is a risk measure such as conditional value at risk, and what does it capture?

Accepted Answer

Conditional value at risk is a risk measure that averages the outcomes in the worst tail of a distribution, such as the worst five percent. Unlike a simple percentile, it reflects how bad losses are beyond the threshold. Optimization models can include it as an objective or constraint to limit downside exposure.

Question 22

What is a discrete event simulation, and what kinds of systems does it model well?

Accepted Answer

Discrete event simulation models a system as a sequence of events that change its state at specific points in time. It is well suited to queues, production lines, call centers, and hospital operations. Analysts use it to compare staffing or layout options before changing a real system.

Question 23

Why is model validation important when a simulation is used to support a decision?

Accepted Answer

Validation checks whether the simulation reproduces the behavior of the real system closely enough to trust its conclusions. Analysts compare outputs with historical records, test extreme inputs, and review the logic with people who know the operation. A polished simulation built on wrong assumptions can produce confident but misleading recommendations.

Question 24

How can a planner use simulation output to compare two candidate decisions fairly?

Accepted Answer

Run both candidate decisions against the same set of random input samples, so that differences in results come from the decisions rather than from luck in the draws. This practice, called common random numbers, reduces noise in the comparison. Report averages along with variability, such as percentiles, before declaring one option better.

Question 25

What is a heuristic, and why do planners use heuristics instead of exact methods?

Accepted Answer

A heuristic is a practical rule or procedure that produces good solutions quickly, without guaranteeing that the solution is optimal. Planners use heuristics when problems are too large or complex for exact methods to finish within available time, or when data are so uncertain that a precise optimum adds little value.

Question 26

How does a greedy algorithm build a solution, and what is its main weakness?

Accepted Answer

A greedy algorithm builds a solution step by step, choosing at each stage the option that looks best immediately. It is fast and easy to implement. Its main weakness is that an early choice can rule out better overall solutions, so the final result may be far from optimal.

Question 27

What is local search, and how does it improve an existing feasible solution?

Accepted Answer

Local search starts from a feasible solution and repeatedly moves to a neighboring solution that is better, based on small changes such as swapping two assignments. It stops when no neighbor improves the objective, which means it has reached a local optimum. Because it can get trapped there, practitioners often add restarts or randomization to explore more of the search space.

Question 28

What is a metaheuristic such as simulated annealing, and what idea allows it to escape poor solutions?

Accepted Answer

A metaheuristic is a general search framework that guides heuristic moves to explore promising regions of a large solution space. Simulated annealing accepts some worse moves with a probability that decreases over time, which lets the search escape local optima early on. As the temperature falls, it becomes more selective.

Question 29

What is a genetic algorithm, and how does it generate new candidate solutions?

Accepted Answer

A genetic algorithm is a population-based search method inspired by biological evolution, where candidate solutions are encoded as chromosomes. New candidates are generated by selecting better parents, combining parts of their encodings through crossover, and introducing small random changes through mutation. Over generations, stronger solutions are kept and weaker ones discarded.

Question 30

What is a tabu search, and why does it keep a memory of recent moves?

Accepted Answer

Tabu search is a local search method that temporarily forbids moves recently made, storing them in a short-term memory called the tabu list. The memory prevents the search from cycling back to solutions it just visited, which encourages exploration beyond the nearest local optimum.

Question 31

When is a heuristic solution considered good enough, and how is its quality usually judged?

Accepted Answer

A heuristic solution is good enough when it meets the decision maker's required performance and fits within practical time and resource limits. Quality is commonly judged by comparing it with a known bound, with the best solution found by other methods, or with benchmark instances where optimal values are known.

Question 32

How does a decomposition approach split a large optimization problem into smaller parts?

Accepted Answer

A decomposition approach divides a large model into subproblems that can be solved separately, often by grouping variables or constraints by region, time period, or product line. The subproblem solutions are coordinated through shared constraints or pricing signals, and the process repeats until the pieces agree.

Question 33

What does it mean to implement a prescriptive recommendation, and why is it more than handing over a number?

Accepted Answer

Implementing a recommendation means embedding the decision logic into the workflow where people or systems act on it. It is more than handing over a number because users must understand the assumptions, trust the inputs, and have authority and time to act. Without these, even a strong model may be ignored.

Question 34

Why should a prescriptive recommendation be compared against a baseline decision before adoption?

Accepted Answer

A baseline, such as the current policy or a simple rule, shows whether the recommended decision actually improves results. Comparing against it reveals the realistic benefit, which may be smaller than the optimized objective suggests once real-world effects are included. Pilot tests or controlled rollouts can provide this comparison with less risk.

Question 35

What is model drift, and why do deployed optimization models need monitoring?

Accepted Answer

Model drift is the growing mismatch that appears when the real environment changes, such as new costs, customer behavior, or supply patterns, so that the model's assumptions no longer match reality. Deployed models therefore need monitoring of input data, constraint status, and actual outcomes compared with predicted ones. Regular recalibration keeps recommendations relevant.

Question 36

How should a team evaluate whether a prescriptive system is delivering value over time?

Accepted Answer

Track a small set of measures tied to the decision's purpose, such as cost saved, service levels met, or revenue gained relative to the baseline. Pair these with diagnostic measures that show how the system is being used. Review the results on a regular schedule so that model, data, and process problems are caught early.

Question 37

What ethical and fairness concerns can arise when a model allocates scarce resources?

Accepted Answer

Optimization models can favor efficiency so strongly that some groups receive systematically less service, even when no one intended that result. Objectives that ignore equity, biased historical data, and constraints that encode unfair rules can all produce unequal outcomes. Teams should review who gains and who bears costs, and document the trade-offs they accepted.

Question 38

Why is human oversight important when a prescriptive system recommends high-stakes actions?

Accepted Answer

Oversight matters because high-stakes actions can have consequences that a model cannot capture, such as safety or legal exposure. Human reviewers can notice context the data missed, question unusual recommendations, and decide when to override the system. Clear escalation rules and records of overrides help the organization learn from both good and bad decisions.

Question 39

What information should an organization document about a model's assumptions and limits?

Accepted Answer

An organization should document the data sources, key assumptions, decision variables, objective, constraints, and known limitations of each model. It should also record who owns the model, when it was last validated, and which decisions it is approved to support. Such documentation supports audits, handoffs between teams, and safe reuse of the model.

Question 40

How do you decide whether a prescriptive model is ready to move from pilot into full use?

Accepted Answer

Readiness depends on evidence that the model performs well on realistic data, that its results beat the baseline, and that users can act on its outputs. Teams should check validation results, sensitivity findings, error handling, and whether monitoring and escalation procedures are in place.

Prescriptive Analytics Systems Flashcards

1Foundations of Prescriptive Analytics

2Optimization Modeling Methods

3Uncertainty and Simulation Methods

4Heuristics and Search Methods

5Deployment, Evaluation, and Governance

Learning Resources