Improve Phase: Solution Design and Optimization

Introduction

The Improve phase converts verified causes of process performance into tested, controlled and economically justified solutions. In DMAIC, the Improve phase follows Measure and Analyze; therefore, its central discipline is not simply to generate ideas, but to design interventions that address validated causes, reduce variation, improve capability and remain feasible in the operating environment.

A technically attractive solution may fail because it is expensive, difficult to implement, unsafe, unacceptable to employees, or vulnerable to new sources of variation. Consequently, improvement requires a systems perspective. The Six Sigma practitioner must combine creativity, structured decision-making, risk analysis, experimentation, simulation and implementation planning.

A useful improvement logic is:

1. Translate validated root causes into design requirements.
2. Generate a broad range of possible solutions.
3. Screen and prioritize alternatives.
4. Assess risk, feasibility and expected benefit.
5. Pilot the selected solution under controlled conditions.
6. Optimize critical settings using data.
7. Confirm that the improved process satisfies customer and business requirements.
8. Transfer the solution to the Control phase.

Learning Objectives

By the end of this chapter, the reader should be able to:

• distinguish solution generation, selection, piloting and optimization;
• use creativity methods to develop countermeasures to validated causes;
• construct and interpret decision matrices and risk-priority analyses;
• apply mistake-proofing and robust-design principles;
• design practical pilots and select appropriate experimental strategies;
• explain response-surface thinking and the role of design of experiments (DOE);
• evaluate improvement using capability, defect, cost and customer measures;
• identify ethical, operational and governance risks in solution deployment.

1. From Root Cause to Design Requirement

An improvement project should not begin with a preferred solution. It should begin with a clear relationship between the validated cause and the required performance change. For example, if analysis shows that excessive injection-moulding defects are associated with unstable mould temperature, the design requirement may be “maintain mould temperature within a specified operating window,” rather than “buy a new controller.” Several solutions may satisfy that requirement: better sensing, automatic control, preventive maintenance, insulation or revised operating procedures.

A solution should satisfy four conditions:

• Causal relevance: it acts on a verified input or mechanism, not merely on a symptom.
• Performance relevance: it improves a critical-to-quality characteristic or reduces harmful variation.
• Operational feasibility: it can be implemented using available technology, skills and capacity.
• Sustainability: it can be controlled, audited and maintained after the project ends.

The distinction between a corrective action and a solution is important. “Retrain operators” is often a weak response when the actual cause is confusing equipment design or an unstable process. Training may be appropriate, but it should not compensate for poor design when a physical or automated countermeasure is possible.

2. Solution Generation Through Structured Creativity

Creativity in Six Sigma is disciplined exploration. The objective is to expand the solution space before narrowing it through evidence and constraints.

2.1 Divergent and convergent thinking

Divergent thinking generates alternatives without prematurely judging them. Convergent thinking evaluates alternatives against agreed criteria. Mixing the two stages can suppress unconventional but valuable ideas. A cross-functional team should therefore separate idea generation from idea selection.

Useful methods include:

• Brainstorming: rapid generation of ideas under rules such as defer judgment, build on others’ ideas and encourage quantity.
• Brainwriting: participants write ideas independently before group discussion, reducing the influence of hierarchy.
• SCAMPER: substitute, combine, adapt, modify, put to another use, eliminate and reverse.
• Process reversal: ask what would guarantee the defect, then reverse those conditions.
• Analogy and benchmarking: examine how another industry controls a similar mechanism.
• TRIZ-inspired contradiction analysis: seek a design that improves one characteristic without worsening another.
• Morphological analysis: decompose the process into functions and combine alternative means for performing each function.

The team should document the causal link, expected mechanism, assumptions and implementation implications for each idea. This record prevents attractive but causally irrelevant solutions from dominating discussion.

2.2 Quality Function Deployment and customer requirements

When several stakeholders are affected, Quality Function Deployment (QFD) can translate the voice of the customer into technical requirements. Customer requirements are weighted according to importance, and possible design characteristics are assessed for their relationship to those requirements. QFD is especially useful when the improvement changes a product, service interface or customer-facing process.

QFD does not prove that a solution will work. It clarifies priorities and trade-offs; empirical testing is still necessary.

3. Screening and Prioritizing Solutions

3.1 Decision matrices

A decision matrix compares alternatives against explicit criteria. Typical criteria include expected defect reduction, customer impact, implementation cost, cycle-time effect, safety, regulatory compliance, employee acceptance, technical feasibility and speed of deployment.

If criterion weights are normalized so that their sum is 1, a weighted score for alternative j is:

where Sⱼ is the total score of alternative j, wᵢ is the weight of criterion i, rᵢⱼ is the rating of alternative j on criterion i, and m is the number of criteria. Ratings must use a consistent scale, such as 1 to 5. The score is meaningful only when criteria are clearly defined and the same direction of preference is used—for example, a higher rating always means better performance.

A decision matrix is an aid to judgment, not an objective replacement for judgment. Sensitivity analysis should be performed by changing major weights and identifying whether the preferred alternative changes. If a small change reverses the ranking, the decision is fragile and additional evidence is warranted.

3.2 Cost–benefit and payback reasoning

Financial analysis should include one-time implementation cost, recurring operating cost, avoided failure cost and expected benefit. A simple payback period is:

The formula assumes relatively stable annual benefits and does not account for the time value of money. For larger investments, discounted cash-flow analysis should be used. Financial benefits must also be separated from accounting claims: an estimated reduction in defects is not a realized saving unless capacity, material, labour or external cost is actually released or avoided.

4. Risk Analysis Before Piloting

4.1 Failure Modes and Effects Analysis

FMEA examines how a proposed solution might fail. For each failure mode, the team considers its effect, cause, current controls, severity, occurrence and detectability. Traditional risk priority number analysis is often expressed as:

where S, O and D are ordinal ratings for severity, occurrence and detectability. The product is not a probability and should not be interpreted as one. Two failure modes with equal RPNs may have very different consequences. A catastrophic but rare failure may deserve priority over a frequent minor failure. Therefore, organizations increasingly use action thresholds, severity gates and engineering judgment alongside RPN.

A strong FMEA asks:

• Could the solution create a new hazard?
• Could it shift defects to another process step?
• What happens during start-up, shutdown, maintenance and abnormal conditions?
• Can the failure be detected before reaching the customer?
• Does the control depend on a single person, sensor or software component?

4.2 Risk response hierarchy

Risk controls should preferably follow the hierarchy of controls: eliminate the hazard, substitute a safer method, apply engineering controls, establish administrative controls and use personal protective equipment where relevant. In quality improvement, engineering controls and automatic detection are generally more robust than reminders or inspection alone.

5. Mistake-Proofing and Robust Design

5.1 Poka-yoke

Mistake-proofing, or poka-yoke, designs the process so that an error is impossible, immediately visible or unable to progress. Examples include asymmetric connectors that cannot be inserted incorrectly, barcode interlocks, fixtures that permit only one orientation, and software validation that prevents incomplete transactions.

The strongest forms are:

1. Prevention: the error cannot occur.
2. Detection at source: the error is identified immediately.
3. Containment: the error cannot reach the next process or customer.
4. Warning: the operator receives an alert, although the error may still occur.

Inspection alone is weaker because it detects rather than prevents the cause and may itself be inconsistent.

5.2 Robust design

Robust design seeks performance that is relatively insensitive to noise factors such as temperature, material variation, operator differences and equipment ageing. The aim is not necessarily to eliminate every source of noise, which may be uneconomic, but to select settings and designs that reduce sensitivity.

A useful quality-loss perspective is:

where L(y) is the loss associated with performance value y, T is the target, and k converts squared deviation into monetary or utility loss. This model assumes that loss increases symmetrically and continuously as performance moves from target. It is a conceptual model; asymmetric specifications or threshold failures may require another loss function.

6. Piloting the Proposed Solution

A pilot is a controlled, limited implementation used to learn before full deployment. It should have a defined scope, owner, duration, inclusion criteria, baseline, success criteria and rollback plan.

6.1 Pilot design

A good pilot specifies:

• the process segment, product family or customer group included;
• the intervention and what remains unchanged;
• the primary response, such as defect rate or cycle time;
• secondary responses, including safety, cost, employee workload and customer experience;
• data-collection procedures and measurement-system requirements;
• decision rules for continuation, modification or termination.

Where feasible, use a comparison group, stepped-wedge rollout or randomized allocation. A before-and-after comparison without controlling for seasonality, mix, learning or demand changes can falsely attribute improvement to the intervention.

For a defect proportion, the observed estimate is:

where d is the number of defect opportunities or units classified as defective under the chosen definition and n is the corresponding total. The definition of the denominator must remain consistent. A pilot should report uncertainty, not only the point estimate. For large-sample planning, the approximate standard error of a proportion is:

where p is the underlying defect probability, or a planning estimate of it. The approximation is less reliable for very small samples or proportions near 0 or 1; exact or Wilson interval methods may then be preferable.

6.2 Pilot example

A service centre processes 4,000 cases per month and historically has 240 rework cases, giving a rework proportion of 240 ÷ 4,000 = 0.06, or 6%. A proposed electronic validation is piloted for 2,000 comparable cases, producing 70 rework cases, or 3.5%.

The absolute reduction is 6.0% − 3.5% = 2.5 percentage points. The relative reduction is:

This result is promising, but it does not alone establish causality. The team must check case mix, measurement consistency, implementation fidelity, false rejections, processing time and user acceptance.

7. Design of Experiments and Response-Surface Thinking

7.1 Why one-factor-at-a-time testing is limited

Changing one factor while holding all others constant is easy to understand but can miss interactions. If the effect of temperature depends on pressure, testing temperature alone may produce a misleading conclusion. DOE varies factors systematically so that main effects and, where appropriate, interactions can be estimated efficiently.

For a response y and factors x₁, x₂, …, a first-order model is:

where β₀ is the intercept, βᵢ are estimated effects, xᵢ are coded or measured factor values, and ε represents unexplained variation. The model assumes that the selected region is adequately approximated by a plane and that errors are appropriately independent for the design and analysis.

When curvature is important, a second-order response-surface model may be used:

The linear terms describe directional effects, squared terms describe curvature and cross-product terms describe interactions. The model is local: it should not be extrapolated far outside the studied region.

7.2 Selecting an experimental design

• Full factorial designs estimate all specified effects and interactions but may require many runs.
• Fractional factorial designs reduce runs at the cost of confounding some effects; alias structures must be understood.
• Central composite and Box–Behnken designs are commonly used for second-order response-surface modelling.
• Taguchi-style robust-design approaches emphasize factor settings that reduce sensitivity to noise, although analysis should still respect the actual design and error structure.
• Sequential experimentation begins with screening, then focuses resources on influential factors and optimization.

Randomization protects against time-related nuisance effects. Replication estimates pure error and improves precision. Blocking controls known nuisance variation such as batch, shift or machine. Center points can help detect curvature in factorial work.

7.3 Multiple responses and desirability

Real processes often have competing goals: minimize defects and cycle time while maximizing yield and safety. A solution may therefore require constrained optimization rather than maximizing one response. The selected setting must satisfy hard requirements first, then optimize preferred outcomes. Statistical significance is not sufficient; practical significance, prediction uncertainty and operating robustness must also be considered.

8. Simulation and Digital Evaluation

Simulation is valuable when live experimentation is costly, unsafe or disruptive. Discrete-event simulation can evaluate queues, capacity, staffing and flow; Monte Carlo simulation can propagate uncertainty in demand, processing time or failure rates; digital twins can connect operational data to a dynamic process model.

Simulation outputs are conditional on assumptions. A model should be verified—does it correctly implement the intended logic?—and validated—does it adequately represent the real system for the decision being made? Input distributions, correlations, warm-up periods, run length and random seeds should be documented. A simulation that reports precise results from poorly validated assumptions creates false confidence.

9. Measuring Improvement and Confirming the Solution

The Improve team should compare baseline and post-intervention performance using operationally meaningful measures:

• defects per unit and defects per million opportunities, when opportunity definitions are stable;
• first-pass yield and rolled throughput yield;
• cycle time, waiting time and throughput;
• process capability, where data are appropriate and the process is reasonably stable;
• cost of poor quality and realized financial benefit;
• safety, compliance and customer outcomes.

For a stable approximately normal characteristic with specification limits, capability indices are commonly defined as:

and

where USL and LSL are upper and lower specification limits, μ is the process mean and σ is the within-process standard deviation. Cₚ measures potential capability if centred; Cₚₖ also reflects displacement from the target or specification centre. These indices are not appropriate as standalone evidence if the process is unstable, the measurement system is inadequate, data are strongly non-normal without suitable treatment, or specification limits have not been validly established.

Confirmation should include a control plan: critical inputs and outputs, measurement method, sampling frequency, reaction rules, ownership and escalation. The Improve phase is complete only when the new process is ready to be sustained by the Control phase.

10. Illustrative Case: Reducing Laboratory Sample-Rejection Rates

Background

A hospital laboratory experiences delays because blood samples are rejected for insufficient volume, incorrect tube type or labelling errors. The project team confirms that rejection is concentrated in emergency-department collections and varies by shift.

Problem and analysis

A validated cause analysis identifies three mechanisms: tube selection is difficult under time pressure, labels are sometimes applied away from the patient, and a pneumatic-tube transport route increases haemolysis for certain samples. The team rejects the assumption that additional generic training alone will solve the problem.

Solution design

The team generates alternatives: colour-coded collection trays, barcode scanning at bedside, a redesigned collection kit, a visual tube-selection guide, revised transport routing and an electronic hard stop for missing identifiers. A weighted matrix gives greatest weight to patient safety, rejection reduction and regulatory compliance. FMEA identifies a risk that a hard stop could delay urgent treatment if the system fails; therefore, an emergency override with mandatory documentation is designed.

Pilot

For four weeks, two emergency-department shifts use the redesigned kit, bedside barcode verification and the revised transport route, while comparable shifts continue with the current process. Measures include rejection proportion, collection time, haemolysis rate, emergency overrides and staff workload. The pilot is stratified by shift and sample type to prevent a change in case mix from being mistaken for improvement.

Optimization

Data show that the transport route improves haemolysis, but the original tray layout increases collection time. A small factorial experiment varies tray layout and guide location. The preferred configuration reduces rejection while keeping median collection time within the service requirement. A response-surface analysis is unnecessary because only discrete design alternatives and one key route factor are being compared; using a complex method would add cost without improving the decision.

Outcome and control transfer

The final design combines the revised transport route, simplified tray layout, bedside identification and a controlled emergency override. The control plan monitors rejection by cause, haemolysis by route, override use and equipment availability. Monthly review is assigned to laboratory quality management, and the process owner receives a reaction plan for deterioration.

Learning points

The case illustrates that improvement is a portfolio of coordinated design decisions rather than a single intervention. It also demonstrates the importance of separating causal analysis from solution preference, balancing quality with flow, considering failure modes created by automation, and selecting analytical methods proportionate to the decision.

11. Managerial, Ethical and Governance Considerations

Improvement can redistribute work, alter autonomy and affect access or safety. Employees should not be blamed for defects when the process design creates predictable error. Data used in pilots must respect privacy, confidentiality and applicable health, employment or industry regulations. Automated decision rules should be explainable, auditable and tested for unequal effects across customer or employee groups.

Managers should distinguish genuine savings from workload transfer. A reduction in one department’s defects is not a total-system benefit if it increases downstream inspection or customer effort. Similarly, an AI-supported recommendation should not be treated as causal evidence merely because it identifies a correlation. Human accountability, model monitoring, data governance and change control remain necessary.

12. Contemporary Developments and Best Practice

Modern Improve projects increasingly combine process mining, real-time sensors, computer vision, digital twins and advanced analytics. These technologies can shorten the feedback loop between process behaviour and intervention. However, technology should follow the validated problem. A sophisticated platform does not compensate for unclear requirements, poor measurement or weak process ownership.

Best practice includes:

• begin with validated causes and customer requirements;
• involve operators, maintainers, compliance specialists and affected customers;
• use a broad idea-generation stage followed by transparent criteria;
• assess total-system effects, not only local metrics;
• use FMEA before implementation and update it after learning;
• prefer prevention and engineered controls over reminders;
• pilot with comparison, stratification or randomization where practical;
• use DOE when interactions and curvature matter;
• report uncertainty and practical significance;
• document assumptions, decision rights and rollback conditions;
• transfer verified controls to an accountable process owner.

Challenges and Limitations

The Improve phase faces several recurring difficulties. Teams may jump to solutions before confirming causes, overvalue senior opinions in decision matrices, or select an intervention because it is easy rather than effective. Pilot results may be distorted by the Hawthorne effect, short observation periods, seasonal demand or changes in product mix. DOE can fail through inadequate factor ranges, confounded effects, poor randomization or insufficient replication. Simulation can mislead when the model is not validated. Finally, capability indices and financial projections can create an appearance of precision unsupported by stable data or credible assumptions.

Professional judgment requires recognizing these limitations, making uncertainty visible and escalating safety or compliance concerns rather than hiding them behind a favorable score.

Chapter Summary

The Improve phase transforms validated causes into tested and sustainable process changes. Effective improvement begins with design requirements, uses structured creativity to generate alternatives, and applies decision matrices, cost reasoning and risk analysis to select promising solutions. Mistake-proofing and robust design reduce dependence on inspection and human memory. Pilots provide controlled learning, while DOE and response-surface methods reveal influential factors, interactions and curvature when optimization is needed. Simulation supports decisions where live experimentation is costly, but its assumptions must be verified and validated. Improvement must be confirmed using appropriate quality, capability, flow, financial and customer measures before transfer to Control. Ethical governance requires attention to safety, privacy, fairness, workload and accountability.

Key Terms

DMAIC: Define, Measure, Analyze, Improve and Control improvement framework.

Design requirement: A measurable condition that a proposed solution must satisfy.

Decision matrix: Structured comparison of alternatives against weighted criteria.

Sensitivity analysis: Examination of how conclusions change when assumptions or weights change.

FMEA: Failure Modes and Effects Analysis used to anticipate and prioritize risks.

RPN: Product of severity, occurrence and detectability ratings in traditional FMEA.

Poka-yoke: Mistake-proofing that prevents or detects errors at source.

Robust design: Design that reduces sensitivity to uncontrollable noise factors.

Pilot: Limited implementation conducted to learn before full deployment.

Response variable: Measured outcome affected by process factors.

Factor: An input variable deliberately varied in an experiment.

Interaction: Condition in which the effect of one factor depends on another factor.

DOE: Design of Experiments for efficient, structured investigation of factors.

Response surface: Estimated relationship between responses and quantitative factors.

Randomization: Allocation or run ordering method that reduces systematic bias.

Blocking: Grouping experimental runs to control known nuisance variation.

Capability: Degree to which a stable process fits within specification limits.

Simulation: Model-based representation used to evaluate process behaviour.

Control plan: Documented method for sustaining performance and responding to deviation.

Cost of poor quality: Costs arising from prevention failures, appraisal, internal failures and external failures.

Review Questions

Short Answer Questions

1. Why should solution design begin with validated causes rather than preferred technologies?
2. What is the difference between prevention and detection in mistake-proofing?
3. Define an interaction effect in DOE.
4. Why is an RPN not a probability of failure?
5. State two conditions under which capability indices may be misleading.

Long Answer Questions

1. Explain how a weighted decision matrix should be constructed, interpreted and tested for sensitivity.
2. Discuss the role of FMEA in designing and piloting a Six Sigma solution.
3. Compare full factorial, fractional factorial and response-surface designs.
4. Explain how a pilot can be designed to strengthen causal conclusions.
5. Discuss the managerial and ethical risks of implementing automated improvement solutions.

Analytical / Application Questions

1. A process has a baseline defect rate of 8% and a pilot defect rate of 5% from comparable volumes. Calculate the absolute and relative reductions, and identify additional evidence needed before rollout.
2. Design a decision matrix for selecting among three solutions to a delayed-service problem. Specify at least five criteria, assign weights and explain how you would test ranking sensitivity.
3. A manufacturing process has two suspected factors and evidence of curvature. Recommend an experimental strategy, including randomization, replication, centre points or second-order modelling, and justify your choices.

Multiple Choice Questions with Answers

Question 1

What is the primary purpose of the Improve phase?

A. To collect the initial voice of the customer

B. To convert validated causes into tested and sustainable solutions

C. To replace measurement-system analysis

D. To close the project before control planning

Correct Answer: B. To convert validated causes into tested and sustainable solutions

Question 2

In a weighted decision matrix, what must the criterion weights usually sum to when normalized?

A. 0

B. 0.5

C. 1

D. 1000

Correct Answer: C. 1

Question 3

Which form of mistake-proofing is generally strongest?

A. A warning poster

B. End-of-line inspection

C. Retraining after every error

D. Designing the error so it cannot occur

Correct Answer: D. Designing the error so it cannot occur

Question 4

Traditional FMEA risk priority number is calculated as:

A. S + O + D

B. S × O × D

C. S ÷ O × D

D. O − S − D

Correct Answer: B. S × O × D

Question 5

A pilot observes 70 rework cases in 2,000 cases. What is the observed rework proportion?

A. 0.35%

B. 2.0%

C. 3.5%

D. 35%

Correct Answer: C. 3.5%

Question 6

Which DOE feature helps protect against time-related nuisance effects?

A. Randomization

B. Removing replication

C. Using one factor only

D. Ignoring batch information

Correct Answer: A. Randomization

Question 7

In a second-order response-surface model, squared terms primarily represent:

A. Randomization

B. Curvature

C. Specification limits

D. Measurement units

Correct Answer: B. Curvature

Question 8

What does Cpk consider that Cp does not?

A. The number of experimental factors

B. The process mean’s position relative to specification limits

C. The financial payback period

D. The sample identification method

Correct Answer: B. The process mean’s position relative to specification limits

Question 9

Why must a simulation model be validated?

A. To ensure its representation is adequate for the intended decision

B. To guarantee that every future outcome is exact

C. To eliminate the need for operational data

D. To make random variation disappear

Correct Answer: A. To ensure its representation is adequate for the intended decision

Question 10

Which statement best describes a pilot?

A. A permanent control plan

B. A limited implementation used to learn before full deployment

C. A replacement for root-cause analysis

D. A financial audit only

Correct Answer: B. A limited implementation used to learn before full deployment