Algorithms

Algorithms define how FinchGE applies evolutionary operators to a population. They coordinate selection, crossover, mutation, evaluation, sorting, and replacement for each generation.

In FinchGE, algorithms are lower-level evolutionary components. A full Grammatical Evolution run is coordinated by the engine, which handles initialization, stopping criteria, checkpointing, logging, and final result construction.

Supported Algorithms

Algorithm	Objective type	Main use
`GeneticAlgorithm`	Single-objective	Standard grammatical evolution with one scalar fitness objective.
`SteadyStateGA`	Single-objective	Genitor-style steady-state replacement; offspring enter the population immediately.
`IslandGA`	Single-objective	Parallel sub-populations with periodic migration to preserve diversity.
`NSGA2`	Multi-objective	Pareto-based optimization using non-dominated sorting and crowding distance.
`NSGA3`	Multi-objective	Many-objective optimization using reference points.

These are the algorithms shipped with FinchGE, not an exhaustive list of what's possible. Any custom evolutionary loop can be added by subclassing BaseAlgorithmSO (single-objective) or BaseAlgorithmMO (multi-objective) and implementing evolve_one_generation.

GeneticAlgorithm

GeneticAlgorithm is the standard single-objective algorithm. It selects parents, applies crossover and mutation, evaluates offspring, sorts individuals by fitness, and applies a replacement strategy to form the next generation.

This is the usual starting point when the experiment has one fitness value, such as error, accuracy, reward, or expression complexity.

Typical operator choices include:

tournament or rank selection
one-point, two-point, or uniform crossover
integer-flip mutation
generational or elitist replacement

SteadyStateGA

SteadyStateGA implements the Genitor model (Whitley, 1989). Instead of building an entire new generation at once, it selects two parents, produces two offspring, evaluates them, and immediately replaces the two worst individuals in the live population. This repeats until population_size offspring have been processed.

Because offspring enter the population immediately, they are eligible as parents or replacement targets later in the same generation. No explicit elitism is needed: the best individuals survive naturally since only the worst individuals are ever displaced.

Use SteadyStateGA when you want finer-grained selection pressure than a generational GA, or when you want to avoid the "generation boundary" effect where all offspring must wait for the whole population to be rebuilt.

IslandGA

IslandGA partitions the population into num_islands independent sub-populations that evolve in isolation using standard GA operators. Every migration_interval generations, the best migration_size individuals from each island migrate to the next island in a ring topology, replacing that island's worst individuals.

Isolation lets different islands explore different regions of the search space; periodic migration shares good solutions across islands to prevent any one of them from stagnating in a local optimum.

population_size must be divisible by num_islands, and migration_size must be smaller than the resulting island size.

Use IslandGA when a single population tends to converge prematurely and you want to trade some convergence speed for sustained diversity.

NSGA-II

NSGA2 is used for multi-objective optimization. Instead of returning a single best individual, it organizes the population into Pareto fronts using non-dominated sorting and uses crowding distance to preserve diversity within each front.

NSGA-II is useful when objectives conflict, for example:

minimizing prediction error while minimizing expression complexity
maximizing reward while minimizing program size
balancing accuracy and interpretability

Use NSGA-II with multi-objective fitness functions and NSGA-aware selection or replacement operators.

NSGA-III

NSGA3 is designed for many-objective optimization. It uses reference points to help maintain diversity across the objective space when crowding distance alone becomes less effective.

NSGA-III is most useful when experiments have more than two objectives or when the search needs stronger coverage across a many-objective Pareto front.

Algorithm and Engine Responsibilities

The algorithm evolves one generation at a time. The engine is responsible for the wider experiment workflow.

Responsibility	Algorithm	Engine
Apply selection, crossover, mutation, and replacement	Yes	No
Evaluate offspring during evolution	Yes	No
Initialize the first population	No	Yes
Manage generation loop and stopping criteria	No	Yes
Handle checkpointing and resuming	No	Yes
Build final experiment result	No	Yes

This split keeps algorithms reusable. The same algorithm can be used in different experiment workflows as long as the grammar, mapper, fitness evaluator, and operators are compatible.

Choosing an Algorithm

Use GeneticAlgorithm when there is one clear optimization target. Use SteadyStateGA when you want continuous, fine-grained replacement instead of discrete generations. Use IslandGA when premature convergence is a concern and you want isolated sub-populations with periodic migration. Use NSGA2 when there are two or more objectives and a Pareto front is desired. Use NSGA3 when the experiment has many objectives and reference-point diversity is important.

For multi-objective algorithms, the result should be interpreted as a Pareto set rather than a single best solution. Selecting one final individual from that set is a modeling decision and should be documented as part of the experiment.