Skip to content

Searcher

The Searcher is a core VerifIA component that explores the input subspace defined by each original‑seed × rule pair. It uses population‑based evolutionary algorithms to both discover challenging derived inputs and quantify overall consistency.

1. Objective Function

  • Goal: Find derived inputs that maximize deviation between the model’s prediction and the rule‑based expected response.
  • Why not a single optimum? A single “worst” violation doesn’t prove inconsistency. Instead, multiple inconsistent points must be observed to statistically reject a model.

2. Statistical Consistency Measurement

  1. Population of derived inputs
    • Initialize a set of candidate points within the constrained subspace.
  2. Evolutionary optimization
    • Evolve the population to increase prediction deviation from the expected output direction (e.g. if rule says “output ↑”, seek predictions with stalled or decreased values).
  3. Consistency ratio
    • For each seed‑rule pair, compute the fraction of derived inputs that violate the rule.
    • Aggregate ratios across all seeds and rules to measure model consistency against the full domain.

3. Population‑Based Evolutionary Algorithms

VerifIA currently supports several nature‑inspired population-based evolutionary metaheuristics:

Algorithm Enum Key Description
Random Sampler RS Fresh random sampling each iteration.
Firefly Algorithm FFA Bioluminescent attraction & random moves.
params: beta_min, gamma, alpha, theta
Genetic Algorithm GA Crossover & mutation on elites/parents.
params: mutation_prob, crossover_prob, elite_ratio
Particle Swarm PSO Velocity & position updates via personal/global best.
params: c_1, c_2, w_max, w_min.
Grey Wolf Optimizer GWO Leadership hierarchy & encircling prey principles.
Moth–Flame Optimizer MFO Spiral movement around ranked “flames”.
params: b
Multi‑Verse Optimizer MVO Universe inflation/deflation & wormhole tunneling.
params: WEP_min, WEP_max, TDR_p.
Salp Swarm Algorithm SSA Chain‑structured population with leader & followers.
Whale Optimization WOA Encircling & spiral bubble-net attacks.
params: b

The algorithm‑specific params control exploration/exploitation balance.

4. Run-wise Configurable Parameters

All population-based searching algorithms accept these two parameters at runtime:

Parameter Purpose Default
pop_size Number of candidates per generation user‑defined
max_iters Iteration budget user‑defined

Tune pop_size & max_iters to trade off coverage/search scope vs runtime/compute budget.

5. Workflow Summary

  1. For each original seed × rule pair:
    a. Define the constrained subspace (white-noise + variation limits + constraints).
    b. Initialize a random population within that subspace.
    c. Evolve the population using the selected algorithm and objective function.
    d. Collect derived inputs that violate the rule.

  2. Compute consistency ratios:

    • ratio = violations_count / total_candidates

  3. Aggregate across all rules and seeds to assess overall domain consistency.

6. Selecting your Searcher

Before running verification, you must seect which searcher VerifIA will use and its parameters:

from verifia.verification import RuleConsistencyVerifier

verifier = RuleConsistencyVerifier("domain_rules.yaml")

# Provide algorithm name and optional parameters
verifier = verifier.using(
    search_algo="PSO",
    search_params={
        "c_1": 0.5,    # lower personal influence
        "c_2": 0.3,    # higher social influence
        "w_max": 0.8,  # reduced max inertia
        "w_min": 0.2   # increased exploitation
    }
)

Failure

Do not include pop_size or max_iters in search_params; they belong in the top‑level .run() call.