liveBook · Manning

Pareto frontier

As explained in Self-Improving Agents

A Pareto front is the set of non-dominated candidates in a multi-objective search, where improving one objective requires worsening another.

Suppose several agent prompts work well, but the higher-quality ones cost more to run. Keeping only one “best” prompt hides a useful choice. The Pareto frontier, or Pareto front, keeps the candidates that offer distinct trade-offs instead. section 4.2.3

Find the front with a pairwise test:

  • Candidate X dominates Y if X is at least as good on every objective and strictly better on at least one. section 4.2.3
  • Drop each candidate dominated by another. Keep every candidate that remains; none can be replaced without giving something up. section 4.2.3

Quality versus cost

Higher quality is better; lower cost is better. A scores 0.90 at cost 9, B scores 0.80 at cost 4, and C scores 0.60 at cost 2. All three stay: moving toward higher quality means paying more. D (0.70, 9) falls to B, which is both better and cheaper. E (0.50, 5) falls to C for the same reason. The survivors form a staircase of choices, not a ranking with one winner. section 4.2.3

Listing 4.4 The domination test and the front (02_pareto_demo.py)
# (name, quality: higher is better, cost: lower is better)
CANDIDATES = [("A", 0.90, 9), ("B", 0.80, 4), ("C", 0.60, 2),
              ("D", 0.70, 9), ("E", 0.50, 5)]
 
def dominates(x, y) -> bool:                             #A
    at_least_as_good = x[1] >= y[1] and x[2] <= y[2]
    strictly_better  = x[1] >  y[1] or  x[2] <  y[2]
    return at_least_as_good and strictly_better
 
def pareto_front(candidates):                            #B
    return [c for c in candidates
            if not any(dominates(o, c) for o in candidates
                       if o is not c)]

The Python listing implements the two-part domination test, then retains candidates for which no other candidate passes it. section 4.2.3

Figure 4.6 The front is the frontier of best trade-offs. Five candidates are scored on quality (up) and cost (across). A, B, and C sit on the front, each the best at some trade-off, so nothing beats them on both and all three are kept; D and E fall in the shaded dominated region, beaten on both axes at once, and are dropped.

The diagram makes the staircase visible: each surviving point offers a trade-off, while dominated points sit behind the frontier. section 4.2.3

In a population-based prompt search, keeping the whole front preserves alternatives for later variation rather than carrying forward only one winner. The distinction matters only when there are multiple objectives: with a single score, the front collapses to a ranked list. A practical next check is to add a candidate that dominates an existing survivor, then one that joins the staircase without displacing anyone. section 4.2.1 section 4.2.3 section 4.5

How Self-Improving Agents covers it

How deep it goes
Defines domination, works through five candidates, and implements a pairwise test; notes that a single-score front collapses to ranking.
Examples
Tests five quality–cost candidates in Python, then asks for a challenger that dominates a survivor and a candidate that joins the front.
What it assumes
Assumes scored candidates and basic Python for the exercise, while defining the domination test before applying it.
What you can do after
Build it from scratch: compare scored candidates, discard dominated ones, and retain the remaining trade-offs for selection.
How it connects
Links multi-objective selection to evolutionary search, prompt populations, and the diversity lost by single-candidate hill climbing.

Also explained in

Compare how 5 books explain Pareto frontier →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Self-Improving Agents. The section links point to the parts of the book it draws on.