Multi-Objective Optimization and Pareto Frontiers
Real decisions rarely have one uncontested objective. Multi-objective optimization exposes tradeoffs; it does not decide which stakeholder values should prevail.
Prerequisites
Complete sensitivity and scenario analysis.
Learning goals
- Define dominance and Pareto efficiency for two maximization objectives.
- Generate a small frontier with the epsilon-constraint method.
- Distinguish a frontier from a recommended policy.
- Explain the limits of weighted sums, scaling, and incomplete stakeholder objectives.
Dominance before preference
Compare predicted total benefit with clinic access . Allocation dominates when it is at least as good on both objectives and strictly better on at least one. A Pareto-dominated point should not be selected under these two objectives. A Pareto-efficient point has no dominating feasible alternative, but many efficient points can remain.
The epsilon-constraint method maximizes benefit while imposing . For clinic minimums from 20 through 25, the frontier trades ten predicted benefit points for each additional clinic batch. Choosing a point still requires stakeholder deliberation about access, measurement, distribution, and omitted harms.
A weighted sum can explore supported points of a convex frontier, but weights mix units and encode a value judgment. It can miss non-supported efficient points in non-convex problems. Reporting only a weighted score hides both issues.
Worked example
import numpy as np
from scipy.optimize import linprog
benefit = np.array([8.0, 6.0])
resources = np.array([[2.0, 1.0], [1.0, 2.0]])
capacity = np.array([40.0, 50.0])
minimums = list(range(20, 26))
frontier = []
for clinic_minimum in minimums:
A = np.vstack([resources, [0.0, -1.0]])
b = np.append(capacity, -clinic_minimum)
result = linprog(-benefit, A_ub=A, b_ub=b,
bounds=[(0.0, None)] * 2, method="highs")
if result.status != 0:
raise RuntimeError(result.message)
frontier.append((clinic_minimum, round(float(-result.fun), 1)))
print(f"clinic_minimums={minimums}")
print(f"pareto_points={frontier}")
print("objectives=predicted_benefit_points,clinic_batches stakeholder_choice=not_solved")
Expected output:
clinic_minimums=[20, 21, 22, 23, 24, 25]
pareto_points=[(20, 200.0), (21, 190.0), (22, 180.0), (23, 170.0), (24, 160.0), (25, 150.0)]
objectives=predicted_benefit_points,clinic_batches stakeholder_choice=not_solved
Each pair reports (minimum clinic batches, maximum predicted benefit points). The threshold is satisfied exactly along this segment. The frontier is conditional on the same continuous model, point estimates, and hard capacities as earlier lessons.
Limitations and common mistakes
- Pareto efficiency is relative to the objectives and constraints included; omitted stakeholders remain omitted.
- More clinic batches are only a defensible objective if access and service quality are meaningfully represented.
- Normalize or retain units when presenting axes; visual distance can otherwise mislead.
- A dense frontier is not a confidence band and does not express outcome uncertainty.
- Selecting a compromise by visual appeal, an arbitrary knee, or hidden weights is still a policy choice.
Practice
Ask two stakeholders to choose a clinic minimum before seeing the solution. Record their reasons, show the frontier, and let them revise. Preserve disagreement and identify one missing objective rather than forcing consensus into a single score.
Next steps
Ticket 59 continues with indivisible choices in integer and binary optimization, then builds toward assignment, flow, scheduling, routing, constraint programming, and nonlinear boundaries.
Primary sources
Checked 2026-10-18: Boyd and Vandenberghe, Convex Optimization, vector optimization; Stanford EE364A, multi-criterion optimization notes; SciPy 1.18.1, linprog and linprog(method="highs").