Choosing Objectives & Measurements
How you define your objectives determines how the optimizer searches for solutions. This article covers when to use each multi-objective strategy and how to handle derived quantities without wasting model capacity.
When to Use Each Multi-Objective Strategy
When you have multiple objectives, choose how to combine them. The choice affects both the quality of results and the computational cost. See the Multi-Objective Optimization article for full details on each strategy.
Default: Use Prioritize (Hierarchy) When Any Objective Is More Important
In most cases, one objective matters more than the others. Use Prioritize (Hierarchy): rank your objectives by importance and set a tolerance for each. The optimizer focuses on the top-priority objective first, then optimizes the next one within the tolerance of the first, and so on.
This is the recommended strategy whenever you can rank your objectives. It converges faster and avoids wasting experiments in regions where a secondary objective is good but the primary one is poor.
Use Pareto Only When All Objectives Are Equally Important
Use Pareto only when your objectives are genuinely equally important and you want to explore the full trade-off front between them. Pareto returns a set of trade-off points rather than a single best solution.
Keep in mind:
Pareto uses hypervolume-based acquisition, which is more computationally expensive.
It works best with 2 to 3 objectives. More objectives (hard limit: 6) make the front harder to approximate and interpret.
If after exploring the trade-off you realize one objective matters more, switch to Prioritize (Hierarchy).
Strategy | When to use | You set |
Prioritize (Hierarchy) | Any objective is more important than the others (most common case) | Priority order + tolerance % |
Pareto | All objectives are equally important and you want to explore trade-offs | Nothing — the optimizer maps the trade-off front |
Weighted Sum | You have a fixed numerical tradeoff — exact relative weights are known | Weight % per objective |
Handling Cost and Derived Quantities
Cost as an objective: If cost matters in your optimization, the recommended approach is to include it as a multi-objective setting using Hierarchy (prioritize). Set your main performance metric (e.g. yield, selectivity) as the top priority, and cost as a secondary objective to minimize. This prevents the optimizer from exploring regions where cost is low but performance is also very low — it first ensures good performance, then optimizes cost within that region.
Derived quantities: Quantities that can be computed deterministically from parameters or other measurements — such as ratios, combined scores, or efficiency metrics — should not be added as variables. If they have a bound that must be met (e.g. a budget limit or mixture sum), add them as a constraint — particularly using a linear constraint if the relationship is linear. Otherwise, add them as a result without objective for tracking purposes.
When to Use a Constraint
Constraints can only be used when the relationship between variables is linear — for example, a weighted sum of variables that must equal a value, stay below a budget, or fall within a range. The platform does not currently support constraints on non-linear derived quantities (like impurity or efficiency ratios).
Examples of what works as a constraint:
Component fractions sum to 100% — a linear sum of variables
Reagent cost ≤ budget — a weighted linear sum of volumes
See the Constraints Overview for how to set these up.
When to Use a Result Without Objective
For derived quantities that are not linear functions of variables (like impurity, efficiency ratios, or combined scores), add them as a result without objective. This lets you track them across iterations without the optimizer trying to learn or predict them. There is currently no way to restrict the search space by non-linear derived quantities.
Examples:
Impurity level — tracked for analysis but not constrainable (not a linear function of variables)
Efficiency ratio — yield per unit cost, recorded for comparison
Combined score — a composite metric you want to monitor
Decision Guide
Is cost important to your optimization?
Yes — add it as a secondary objective using Hierarchy, with your main performance metric as the top priority.
Is the quantity a real experimental outcome (not computable from other values)?
Yes — add it as an objective (Maximize, Minimize, or Target).
No — it is derived. Go to the next question.
Is it a linear function of your variables with a bound that must be met?
Yes — add it as a linear constraint.
No — add it as a result without objective for tracking.