Exploration vs. exploitation

The central tension of metaheuristic-optimization: explore the search space widely (find new, possibly better regions; avoid getting stuck in local optima) vs. exploit the best-known region (refine it toward the optimum). Too much exploration wanders; too much exploitation converges prematurely. Every optimizer is, in effect, a different schedule for trading the two off.

How the founding four resolve it

• exchange-market-algorithm — makes it explicit: a balanced-market phase (conservative, exploit via imitation of elites) and a fluctuating-market phase (aggressive explore: random resets + opposition-based learning).
• backtracking-search-algorithm — injects exploration via a shuffled historical population (movement relative to past positions), then exploits with greedy selection.
• cma-es — shapes exploration: the adapted covariance matrix stretches the sampling distribution along promising directions; step-size control modulates breadth.
• deterministic-oscillatory-search — explores by oscillating (bounce + halve velocity on worsening fitness), then exploits by swarming toward the global best when oscillation stalls.

And the wider corpus

• particle-swarm-optimization — momentum (w) plus pull to personal + global best; prone to collapsing into exploitation (premature convergence).
• grey-wolf-optimizer — the coefficient a decays 2 → 0, scheduling a hard shift from exploration (search) to exploitation (attack).
• ant-colony-optimization — pheromone reinforcement exploits; evaporation restores exploration.
• artificial-bee-colony — onlookers exploit good sources; scouts abandon stalled ones to explore.
• evolution-strategies — the clearest knob: (μ+λ) keeps parents (more exploration) vs. (μ,λ) forces turnover (faster, risk of premature convergence).
• genetic-algorithm / differential-evolution — explore via mutation + crossover, exploit via (greedy) selection; GA fights premature convergence with niche penalties / random immigrants, DE self-scales its step from the population spread.
• simulated-annealing — makes the balance temporal: high temperature explores (accepts uphill moves), cooling shifts it to exploitation.
• bayesian-optimization — the most explicit: the acquisition function numerically trades the two off, sampling where expected improvement or model uncertainty is highest.

Related

metaheuristic-optimization · no-free-lunch-theorem · cma-es · exchange-market-algorithm · deterministic-oscillatory-search · evolution-strategies · particle-swarm-optimization · grey-wolf-optimizer · genetic-algorithm · differential-evolution · simulated-annealing · bayesian-optimization