Core concepts
HeuristicLib separates the optimization problem from the search method. Reuse the same problem when testing or comparing algorithm configurations.
The six parts of a run
| Concept | Question it answers | Example |
|---|---|---|
| Candidate | What does one possible answer look like? | A vector of four real numbers |
| Search space | Which candidates are valid? | Four values between -5.12 and 5.12 |
| Problem | How is a candidate evaluated? | Calculate the Rastrigin function |
| Objective | Which result is better? | A smaller function value |
| Algorithm and operators | How do we explore? | Genetic algorithm with crossover and mutation |
| Search state | What has happened so far? | Population and objective values after generation 12 |
Problem
Search space
Evaluation
Objective
Algorithm configuration
Operators
Population and budgets
Random source
Recorded root seed
Algorithm run
Fresh execution state
Search states
Evaluated candidates and progress counters
Candidate and search space
A candidate is one proposed solution. Its type represents the shape of the answer, such as RealVector, IntegerVector, BoolVector or Permutation.
The search space owns validity rules for that type. A RealVectorSearchSpace can specify length and bounds. Operators use that information when creating or modifying candidates.
Read Search spaces for the built-in choices.
Problem and objective
A problem combines three facts:
- the search space
- the evaluation of a candidate
- the direction and structure of the objective
Evaluation produces an objective vector, even for a single objective problem. The objective then defines whether lower or higher values are better and how results are compared.
Read Problems and Objective vectors.
Algorithm and operators
An algorithm owns the search process and its stopping budget. Operators perform focused steps within that process. A genetic algorithm usually needs operators to create candidates, select parents, cross parents and mutate offspring.
This split lets you change the variation strategy without implementing another generation loop. Type parameters and search space checks catch incompatible combinations early.
Read Algorithms and Operators.
Search states
Algorithms yield immutable snapshots of useful progress. A population based state contains evaluated candidates. An iterative state also reports counters such as completed iterations.
You can consume every state for progress reporting and analysis or ask only for the final state. Read Search states and Running algorithms.
Randomness is an input
Stochastic algorithms receive an explicit random number generator. A recorded seed is part of an experiment configuration, just like the population size or mutation rate. It makes a surprising run reproducible.
Read Reproducible randomness.
Important guarantees
- Search spaces define candidate validity.
- Problems own evaluation and objective semantics.
- Evaluated candidates keep a candidate together with its objective vector.
- Each call to
StreamorCompleteAsyncstarts a fresh run. - The same configuration and seed are intended to produce the same stochastic sequence.
- One run is evidence, not a reliable comparison of stochastic algorithms.