Predictive Pareto Planning
A single execution plan optimizes for one objective. Pareto planning generates the full set of non-dominated trade-offs across four objectives simultaneously, letting you choose the plan that best fits your mission constraints.
How it works — CAE evaluates thousands of candidate plans with different relaxation combinations, filters dominated solutions, and returns only the Pareto-optimal frontier.
Objectives
Every plan is scored on four axes:
| Objective | Unit | Direction | Description |
|---|
| Latency | seconds | minimize | Total time from job start to final output delivery |
| Energy | watt-hours | minimize | Total on-board energy consumed across all steps |
| Reliability | probability | maximize | Delivery confidence accounting for link errors and retransmissions |
| Quality | 0-1 score | maximize | Output fidelity — driven by quality tier selection in window steps |
Relaxation Types
To explore the frontier, the planner applies controlled relaxations — each trades one objective for gains in others.
| Relaxation | Trades Away | Gains |
|---|
extend_windows | Latency | Reliability, Quality — more time allows higher tiers and more retransmission margin |
skip_checkpoint | Reliability | Latency, Energy — removing checkpoints saves time and power |
reduce_fec | Reliability | Latency, Energy — less FEC overhead means smaller transfers |
lower_quality_tier | Quality | Latency, Energy — lower tiers complete faster with less compute |
Dominance Filtering
Given two plans A and B, A dominates B if A is at least as good as B on all four objectives and strictly better on at least one. The Pareto frontier is the set of plans that no other plan dominates.
Plan A: latency=4200s energy=18Wh reliability=0.97 quality=0.85
Plan B: latency=5100s energy=22Wh reliability=0.95 quality=0.80
→ A dominates B (better on all four axes)
Plan C: latency=3800s energy=24Wh reliability=0.93 quality=0.90
→ A does not dominate C (C has better latency and quality)
→ Both A and C are on the frontier
API Usage
Single-Satellite Pareto
Add pareto: true to a standard plan request:
curl -X POST https://rotastellar-cae.subhadip-mitra.workers.dev/v1/plan \
-H "Content-Type: application/json" \
-H "Origin: https://rotastellar.com" \
-d '{
"satellite_id": "25544",
"preset_id": "onboard-ml-inference",
"pareto": true
}'
frontier array instead of a single plan:
{
"id": "pareto-a92f33e8-...",
"frontier": [
{
"plan_index": 0,
"objectives": {
"latency_s": 3240,
"energy_wh": 14.2,
"reliability": 0.991,
"quality": 0.95
},
"relaxations_applied": [],
"plan": { ... }
},
{
"plan_index": 1,
"objectives": {
"latency_s": 2880,
"energy_wh": 12.8,
"reliability": 0.967,
"quality": 0.85
},
"relaxations_applied": ["lower_quality_tier", "reduce_fec"],
"plan": { ... }
}
],
"frontier_size": 5,
"candidates_evaluated": 128,
"dominated_filtered": 123
}
Fleet-Level Pareto
For constellation workloads, use the fleet Pareto endpoint:
curl -X POST https://rotastellar-cae.subhadip-mitra.workers.dev/v1/constellation/pareto \
-H "Content-Type: application/json" \
-H "Origin: https://rotastellar.com" \
-d '{
"satellite_ids": ["25544", "48274", "55909"],
"preset_id": "split-learning",
"pareto": true
}'
Response Parameters
| Field | Type | Description |
|---|
frontier | array | Array of Pareto-optimal plan variants |
frontier[].plan_index | number | Index within the frontier (0 = baseline) |
frontier[].objectives | object | Objective scores for this variant |
frontier[].relaxations_applied | array | Which relaxations produced this variant |
frontier[].plan | object | Full plan object (same schema as standard plans) |
frontier_size | number | Number of non-dominated solutions |
candidates_evaluated | number | Total candidate plans generated |
dominated_filtered | number | Candidates eliminated by dominance filtering |
Console Integration
In the RotaStellar Console, the Generate Trade-offs button on the plan detail page triggers a Pareto analysis. Results are displayed as an interactive scatter chart where:
- Each axis maps to one of the four objectives
- Each point is a Pareto-optimal plan variant
- Clicking a point loads the full plan detail
- Hovering shows the relaxations applied and objective scores
Pareto planning takes longer than single-objective planning because the planner must evaluate and filter many candidates. For complex workloads with many steps, expect 2-5x the normal planning time.
