Population Based Search Orderings

Input: $ARGUMENTS

Overview

Instead of pursuing one solution, maintain a population of candidates. Evaluate, select, recombine, and iterate. Finds better solutions than single-track search when the solution space is large and deceptive.

Ordering Rules

Rule 1: Parallel Evaluation

• Evaluate all candidates in the population before selecting
• Don’t commit to one path prematurely
• When: multiple plausible approaches, uncertain which is best

Rule 2: Select and Recombine

• Keep the best candidates, combine their strengths
• “Take the best parts of each approach” — not metaphorical, literal
• When: solutions have modular components that can be mixed

Rule 3: Iterative Deepening

• Search to depth 1, then depth 2, then depth 3…
• Guarantees finding the shallowest solution while managing memory
• When: unknown solution depth, memory constraints

Rule 4: Beam Search — Top-K

• At each step, keep only the K best candidates
• Expand each, evaluate, keep top K again
• When: space too large for full exploration, need approximate best

Rule 5: Mutation — Random Perturbation

• After selecting best candidates, randomly modify some
• Prevents premature convergence on local optima
• When: stuck, all candidates too similar

Application

Step 1: Initialize Population

• Generate diverse starting candidates (don’t start all the same)

Step 2: Evaluate All

• Score each candidate on the objective

Step 3: Select, Recombine, Mutate

• Keep top performers, combine, add random variation
• Repeat until convergence or budget exhausted

When to Use

• Complex optimization, creative ideation, strategy selection
• When one approach feels risky

Verification

• Population diverse at initialization
• All candidates evaluated before selection
• Mutation preventing premature convergence
• Best solution tracked across all generations