Skip to main content
Studies · CA Air Quality · Investigation 29 · Phase 3

Multi-objective Pareto frontier for California AQ portfolios

Answers a different question than Inv 23 — not a re-ranking. 3-objective deterministic frontier vs scalar net-benefit MC. NSGA-II (Deb et al. 2002) on a 5-dim portfolio design space surfaces 100 non-dominated points. Of 6 candidate seeds (4 Inv 23, 2 new), 4 are dominated. Building retrofits re-enter the frontier when combined with indoor AQ spending.

100
Pareto-optimal portfolios
4/6
Inv 23 seeds dominated
0.23
Max DAC share on frontier
+133
Deaths over E_smart at same $
The Question

Is building retrofit off-frontier, or just equity-undervalued?

Inv 12 and Inv 23 both optimize a scalar objective. The headline — "wildfire beats transport beats building" — rests on a single dimension of value (deaths avoided). The CEC GFO-25-304 explicitly asks for DAC equity analysis, and Inv 19 showed indoor AQ retrofits take B2 from 47 to 341 deaths avoided (~7× outdoor-only benefit). Three-objective tradeoff: health, equity, cost.

NSGA-II (Deb et al. 2002) solves this by computing the full Pareto frontier — the set of non-dominated portfolios where no single objective can improve without sacrificing another. Each frontier point is a policy option.

Multi-Objective Fidelity Ladder

From single-scalar to full frontier

L1
Expected-value maximize health Single scalar objective; ignores equity and risk. Inv 12 Phase 1 baseline.
1 obj
dimension
L2
Weighted sum (health + equity) Hand-tuned weights; unstable at corners. Inv 02 tract-level reporting.
2 obj
dimension
L3
Constrained optimization Max health subject to DAC >= 25% floor. Useful but hides tradeoffs.
1 + K
constraints
L4
Robust portfolio (CVaR / DRO / info-gap) Inv 23. Single-objective with uncertainty, 5 robustness lenses.
1 obj
robust
L5
NSGA-II multi-objective Pareto Full 3-objective frontier. 80 generations, 100 population, SBX + polynomial mutation.
3 obj
pareto
Pareto Frontier

100 non-dominated portfolios

Three-objective NSGA-II run: 100 population x 80 generations with SBX crossover (η=15) and polynomial mutation (η=20, p=0.2). Design space spans wildfire reduction 0-10%, transport/building spend 0-$4B each, indoor AQ 0-$2B, DTE on/off. Objectives are evaluated through a deterministic linear surrogate (rfaq/optimization/pareto_frontier.py) calibrated to Inv 23 mean deaths/cost for portfolios A, B, C (exact) and E (within ~45 deaths). The surrogate does not re-draw Inv 21's CRF posterior or apply Inv 23's VSL scalarization.

DAC 0.200.24 $0B$2B$4B$6B$8B$10B$12B 1,0001,5002,0002,5003,0003,500 Portfolio cost ($B) Deaths avoided (deaths/yr)

Each dot is a Pareto-optimal portfolio. Color = DAC health share (violet ~0.20, gold ~0.24). X-axis = cost ($B), Y-axis = deaths avoided / yr.

Max-health corner
3853
deaths avoided / yr
$12.97B · DAC 0.229
Max-equity corner
2754
deaths avoided / yr
$5.27B · DAC 0.231
Min-cost corner (free lunch)
1013
deaths avoided / yr
$0.00B · DAC 0.210
Corner Solutions

Three extremes on the frontier

Max-health corner: 3853 deaths/yr at $12.97B. Loads on wildfire reduction and transport: x = [0.1, 3.89, 3.79, 2.0, 0.01].

Max-equity corner: DAC share 0.231 at $5.27B. Weighted toward indoor AQ and building retrofit.

Min-cost corner: Free-lunch baseline. T1+B1+DTE only; 1013 deaths/yr at $0B.

Seed Portfolio Dominance Check

Which seed portfolios does the frontier dominate?

Six candidate seeds: 4 Inv 23 portfolios (A, B, C, E) plus 2 new design points constructed for this study (indoor_focus, balanced_2B). The NSGA-II Pareto set dominates 4 and leaves 2 on or near the frontier. Inv 23's F_maximum ($13.9B) and G_sequential (BO-optimal) are excluded because they use levers outside this 5-dim basis.

PortfolioDeathsDAC shareCostStatus
A_free_lunch10150.210$0.00BOn frontier
B_transport_2B13830.208$2.00BDominated
C_wildfire17390.219$1.65BOn frontier
E_smart17890.220$2.15BDominated
indoor_focus13150.233$2.00BDominated
balanced_2B16660.221$2.49BDominated

Headline: E_smart (wildfire 5% + building $0.5B + DTE) is dominated. The best Pareto improvement delivers +133 deaths/yr with essentially the same cost, by rebalancing toward indoor AQ. Building retrofits are not off-frontier — they belong in a balanced indoor + wildfire portfolio that Inv 23's discrete seeds never enumerate.

Scope of "dominated": deterministic 3-objective frontier — deaths, DAC-deaths, cost. No VSL, no Monte Carlo, no robustness lens.

Inv 23 ranks the same seeds on a scalar net-benefit (deaths × VSL − cost) over MC draws of the Inv 21 CRF posterior, with CVaR and info-gap layers on top. Re-scoring these Pareto points through that pipeline is the natural next step.

Decision implication

Portfolio recommendation

Rather than committing to Inv 23's E_smart, planners should consider a family of Pareto-optimal options with explicit (health, equity, cost) knobs. The max-equity corner reaches DAC share 0.231, essentially matching the 0.233 ceiling set by indoor_focus (the highest-DAC seed in this study's candidate set, itself dominated by the frontier). It does so without sacrificing 71% of the max-health target.

Method: NSGA-II (Deb et al. 2002), SBX + polynomial mutation, 80 generations, 3-objective non-dominated sorting with crowding distance. Objective surrogate: deterministic linear model in rfaq/optimization/pareto_frontier.py — NOT Inv 23's MC net-benefit pipeline.