Description
Fits a quadratic model of an outcome over continuous design parameters to locate optimal settings, moving sequentially toward the optimum.
Many UX parameters are continuous: font size, animation duration, recommendation count, delay before a prompt, price. RSM treats the outcome as a surface over these parameters. It starts with first-order designs and steepest ascent to move toward improvement, then uses second-order designs such as central composite or Box-Behnken to fit curvature and find the optimum or a ridge of near-optimal settings. It also reveals trade-offs among multiple outcomes via desirability functions. This beats guessing a few discrete A/B levels and is efficient in the number of conditions.
Attributes
- Discipline
- Experimentation & Causal Inference
- Loss type
- Unmigrated
- Stage
- Design, Evaluation
- Data
- Quantitative
- Effort
- Weeks
Steps
- Choose 2-4 continuous factors and plausible ranges, and the outcome(s) to optimise.
- Run a first-order design (e.g. 2^k factorial with centre points) and fit a linear model.
- If far from optimum, move along the path of steepest ascent with a few runs.
- Near the optimum, run a central composite or Box-Behnken design and fit a quadratic model.
- Locate the stationary point, check if it is a maximum, minimum or saddle, and examine contour plots.
- Use desirability functions for multiple outcomes and confirm the chosen setting in a validation test.
Sources
- Box & Wilson, On the experimental attainment of optimum conditions (1951), JRSS B
- Myers, Montgomery & Anderson-Cook, Response Surface Methodology (4th ed., 2016)
Also known as
- RSM
- Central composite design
- Box-Behnken design
- Multiresponse optimisation
Categories
- Analytics & Behaviour
- Experimentation & Causal Inference