greedy-selection-algorithm

name	greedy-selection-algorithm
description	Use when building selection algorithms that must balance multiple diversity/coverage constraints (e.g., selecting test cases, sampling datasets, portfolio optimization). Especially useful when hard co...
tags	["code-pattern","learned","algorithms","selection","diversity","constraints","migration"]

Greedy selection algorithm pattern for subset selection with diversity constraints

When to Use

Use when building selection algorithms that must balance multiple diversity/coverage constraints (e.g., selecting test cases, sampling datasets, portfolio optimization). Especially useful when hard constraints can conflict and you need observable, debuggable scoring.

Code Pattern

When implementing a greedy selection algorithm that needs to balance multiple constraints (diversity, coverage, distribution):

1. Tracking Sets Pattern: Maintain separate tracking sets for each constraint dimension

   • Set for selected items (deduplication)
   • Set for categorical attributes (e.g., instructors, tags)
   • Map for distribution tracking (e.g., era counts)

2. Scoring Function: Break down total score into named components

   • Return score breakdown object (aids debugging)
   • Weight factors by importance (uniqueness > coverage > balance)
   • Use bonuses (+10, +5, +1) instead of multiplication (clearer semantics)

3. Superset Guarantee: Always seed with previous results for incremental expansion

   • Initialize tracking from seed data
   • Prevents regression in constraint satisfaction

4. Constraint Validation: Separate scoring from validation

   • Score = selection priority (greedy step)
   • Validation = hard constraints (exit criteria)
   • Report both target and actual values

5. Progress Reporting: Log every Nth iteration with score breakdown

   • Helps debug why selection stalls or diverges
   • Shows which constraints are dominating selection

Example structure:

function scoreItem(item, selected, tracking, constraints) {
  const breakdown = {
    uniqueness: !tracking.has(item.key) ? 10 : 0,
    coverage: item.newAttributes.filter(a => !tracking.seen.has(a)).length,
    balance: isUnderrepresented(item.category) ? 5 : 0,
  };
  return { ...item, score: sum(breakdown), breakdown };
}

while (selected.size < target) {
  const remaining = candidates.filter(c => !selected.has(c));
  const scored = remaining.map(c => scoreItem(c, selected, tracking, constraints));
  scored.sort((a, b) => b.score - a.score);
  
  const winner = scored[0];
  selected.add(winner);
  updateTracking(winner, tracking);
  
  if (iteration % 10 === 0) logProgress(winner);
}

Reference Files

• migration/scripts/select-courses.ts
• migration/scripts/query-phase2-courses.ts

---

Learned from swarm execution on 2025-12-13