Taguchi Robust Design

Description

Chooses design settings that perform well across uncontrollable noise conditions, optimising signal-to-noise ratios rather than average performance.

Taguchi distinguished control factors (designer chooses) from noise factors (users, environments, devices the designer cannot control). Using crossed or combined arrays, experiments vary both, and robustness is measured with signal-to-noise ratios or variance models. The quality loss function quantifies that any deviation from target imposes cost, not just out-of-specification results. For UX, noise factors are everywhere: screen size, lighting, connectivity, literacy, age, assistive tech. Robust design finds layouts and parameters whose performance is stable across them, rather than optimal for the average user and poor for many.

Attributes

Loss type
Unmigrated
Stage
Design
Effort
Weeks

Steps

  1. List control factors (design options) and noise factors (device, bandwidth, ambient light, user skill).
  2. Choose an orthogonal array for control factors and a noise array or deliberate noise sampling.
  3. Run each control combination across noise conditions and record performance.
  4. Compute S/N ratios (e.g. larger-the-better or nominal-the-best) or model mean and variance separately.
  5. Select settings that maximise robustness, then adjust a scaling factor to hit the target mean.
  6. Confirm with a validation run under fresh noise conditions.

Sources

  • Taguchi, Introduction to Quality Engineering (1986)
  • Phadke, Quality Engineering Using Robust Design (1989)
  • Nair (ed.), Taguchi's parameter design: A panel discussion (1992), Technometrics

Also known as

  • Taguchi methods
  • Robust parameter design
  • Quality loss function

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference