User Guide

• Multivariate Analysis
  • When to Use These Methods
  • Available Methods
• Cross-Validation
  • Selecting the Number of Components
  • Element-wise Cross-Validation (PCA)
  • Choosing among the schemes
  • What PRESS is measured in
  • Selection Rules
  • PLS Component Selection
  • Comparing the fitted and the validated R2
  • Validating the latent structure, not only the predictions
  • PLS Beta Coefficient Error Bars
• Model Evaluation and Visualization
  • Choosing the Number of Components
  • Explained Variance
  • Correlation Loadings
  • Observed versus Predicted
  • Regression Coefficients
  • Comparing Two Data Blocks
• Experimental Strategy Recommendation
  • When to Use This Tool
  • Concepts
  • Quick Start
  • Interpreting the Output
  • Working with Budget Constraints
  • Using Prior Knowledge
  • Domain-Specific Strategies
  • Hard-to-Change Factors
  • Multiple Responses
  • See Also
• The Next Round of an Experiment
  • A worked example
  • The loop
  • The rules, in order
  • How the terms are judged
  • How well it works
• Run Sheets
  • Writing a sheet
  • Reading the results back
• Design families and how each one is checked
• The Design Catalog
  • Searching it in Python
  • What a record holds
  • What a design withstands
  • Comparing designs
  • The errata
  • How closely each case reproduces
  • Opening a design, and limits on its settings
  • Proposing a design or a case study
• Evaluating Design Quality
  • When to Use This Tool
  • Quick Start
  • The Metrics
  • Region Sampling and Reproducibility
  • Tunable FDS Curve
  • Comparing Designs on Nine Measures
  • Testing an analysis on a known process
  • Simulating the Studies a Design Would Give
  • See Also
  • Run order and time trends
• Generating OMARS Designs
  • Quick start
  • The method
  • Performance: iterations and timing by factor count
  • Limitations
  • References
• Designs over a Constrained Region
  • Worked example: a shared heat budget
  • How the design is chosen
  • Writing constraints
  • Combining with other options
  • Combinations of levels that cannot be run
  • Judging the design over its own region
  • Choosing the criterion: D, I or A
  • Choosing runs from a list: candidate sets
  • Running the design in blocks
  • Optimising inside the region
  • Constrained mixtures
  • Mixture components with process factors
• Limits on the Settings
  • When a factor’s own range is a limit
  • Writing limits as text
  • What the limits leave
  • Renaming a factor inside a limit
  • Proposing a limit from the settings you cannot run
  • Adding runs inside the limits
  • Pictures of the region
  • Comparing designs over the region
  • Plotting a response over the region
• Supersaturated and Space-Filling Designs
  • Supersaturated designs: more factors than runs
  • Space-filling designs: no model assumed
  • Constrained regions and mixtures
• Split-plot experiments
  • The corrosion experiment
  • From a split-plot design to its analysis
  • How the analysis works
  • References
• Models fitted across blocks
• Counts, proportions and skewed responses: generalized linear models
  • The coupon experiment
  • Families, links and conventions
• Descriptive panel data: validate, check the panel, relate to the product
  • The data contract
  • Designed versus observational
  • Step 0: get the data into long form
  • Step 1: validate
  • Step 2 and 3: check the panel and relate
  • Telling genuine drivers from proxies: the predictive descriptors
  • Correcting the panel: the Mixed Assessor Model
  • Comparing designed treatments: ANOVA, Tukey HSD and Dunnett
  • Worked example
  • Using the tools from an agent
• Panel diagnostics: can this attribute be modelled?
  • An attribute pinned against a bound
  • The response that suits such an attribute
  • Unequal assessor variance, mistaken for scale use
  • An empty panel is an error
  • References
  • API
• Chemistry: preparing a product-by-compound block
  • The order is fixed
  • A zero is not self-describing
  • A trimmed compound is not a discarded compound
  • Choosing a transform
  • detected_only is off for a reason
  • Preprocessing a held-out fold honestly
  • API
• Permutation nulls: is there anything here at all?
  • Two ways that do not work
  • What does work
  • The p-value has a floor
  • Testing the whole procedure, not one model
  • Recovering the expected class
  • API
• Interaction terms (provisional)
  • Does the pair support a term at all?
  • Building the terms
  • How stable is a selection?
  • References
  • API
• The batch (bioreactor) simulator
  • The process model
  • Realism: the sensitivity budget
  • The three claims, in code
  • The golden batch and the adaptation ceiling
  • Agent tools
• Mid-course correction of batch processes
  • The pieces
  • The executed comparison
  • Where to put the decision point
  • Practical notes