Improve: Generate, Select, Test, Implement
What Improve must produce
Improve converts verified causes into changes that move the Y. Concepts: idea generation (brainstorming, benchmarking, reversing the problem) before any selection; a weighted solution matrix or Pugh matrix scored against impact, cost, risk and time; Lean tools that remove non-value-added work - the eight wastes, value stream mapping, 5S, kanban, SMED, standard work, and poka-yoke (error-proofing, preferring prevention over detection); design of experiments when several factors are suspected - factors and levels, full factorial runs = levels^factors, main effects and interactions, randomization to protect against nuisance variables, replication to estimate error, and blocking; a pilot before full rollout; updated FMEA and cost-benefit; the To-Be process map; and an implementation plan with named owners and dates.
Traps questions exploit
- OFAT (one factor at a time) misses interactions - questions love the scenario where two factors only cause trouble together.
- Mixing up main effect and interaction wording; an interaction means the effect of A depends on the level of B.
- Forgetting why randomization exists (it is not about fairness or sample size).
- Assuming a successful pilot guarantees full-scale success - scale brings new variation.
- Choosing a solution before generating alternatives, or using a kaizen event where a measured DMAIC project is required.
Worked example
Two factors, two levels each: 2² = 4 runs. The main effect of A is the average response with A high minus the average with A low, averaged across both levels of B. If that difference is large at B-low and near zero at B-high, you have an interaction - and an OFAT study, which changes one factor while holding the other fixed, would have reported only half the story.
Rule of thumb: pilot small, measure the same CTQ with the same operational definition, and compare against the baseline, not against a feeling.