Local Search#

Local search optimizers explore the neighborhood of the current best solution. They are fast but may get stuck in local optima.

These optimizers work by iteratively improving from a starting point, making small adjustments to find better solutions nearby.

Hill Climbing#

The simplest local search: always move to a better neighbor.

from hyperactive.opt.gfo import HillClimbing

optimizer = HillClimbing(
    search_space=search_space,
    n_iter=5,
    experiment=objective,
)

Simulated Annealing#

Like hill climbing, but sometimes accepts worse solutions to escape local optima. The “temperature” controls exploration vs exploitation.

from hyperactive.opt.gfo import SimulatedAnnealing

optimizer = SimulatedAnnealing(
    search_space=search_space,
    n_iter=5,
    experiment=objective,
)

Stochastic Hill Climbing#

Hill climbing with a probability of accepting worse solutions. The p_accept parameter controls exploration: higher values make it more likely to accept non-improving moves, helping escape local optima.

from hyperactive.opt.gfo import StochasticHillClimbing

optimizer = StochasticHillClimbing(
    search_space=search_space,
    n_iter=5,
    experiment=objective,
    p_accept=0.3,  # Probability of accepting worse solutions
)

Repulsing Hill Climbing#

Remembers visited regions and avoids them, encouraging broader exploration.

from hyperactive.opt.gfo import RepulsingHillClimbing

optimizer = RepulsingHillClimbing(
    search_space=search_space,
    n_iter=5,
    experiment=objective,
)

Downhill Simplex (Nelder-Mead)#

Uses a simplex of points to navigate the search space. Good for continuous problems.

from hyperactive.opt.gfo import DownhillSimplexOptimizer

optimizer = DownhillSimplexOptimizer(
    search_space=search_space,
    n_iter=5,
    experiment=objective,
)

When to Use Local Search#

Local search optimizers are best suited for:

Quick initial exploration: Get a baseline result fast
Smooth, unimodal landscapes: When there’s a single optimum to find
Refinement: Fine-tune solutions found by global methods
Limited budget: When you can only afford a few evaluations

Consider using global or population-based methods if:

Your landscape has many local optima
You’re not finding good solutions with local search
You need more robust exploration