Evolutionary algorithms are stochastic, population-based search methods that iteratively improve candidate solutions using fitness-based evaluation. Typically, this involves some sort of variation/mutation/recombination to generate diversity and explore new regions of the search space, along with selection to favor high-fitness solutions. Thus, evolutionary algorithms do not learn a model explicitly; they learn by accumulating and preserving useful information inside high-fitness solutions.
Fitness
Fitness assigns a scalar value to each candidate solution , measuring how well it satisfies the problem objective and constraints. Fitness induces a preference ordering over the search space, defining what the algorithm considers good, bad, or acceptable. It is determined entirely by the environment (objective function, constraints, performance criteria).
For a geometric interpretation, we can consider the search to be looking for a target region , signifying solutions with acceptable fitness. Every candidate solution corresponds to a point in parameter space; the fitness function induces a landscape.