The problem of finding the best possible solution across the entire search space of an optimization problem, not just a locally optimal solution that is better than its immediate neighbors. Global optimization is the challenge underlying media budget allocation, hyperparameter search, and any AI-driven optimization task where the landscape has many local optima that simpler methods get trapped in.
Also known as global search, global minimum, global maximum search
Global optimization seeks the input configuration that minimizes or maximizes an objective function across its entire feasible domain, as opposed to local optimization, which finds a configuration better than all nearby configurations but potentially far from the true global best. The distinction matters when the objective function is non-convex: a loss function with multiple valleys means gradient descent will converge to the nearest valley from its starting point rather than the globally deepest one. The global minimum may be orders of magnitude better than the local minimum the algorithm found, but gradient-based methods have no mechanism for escaping a local optimum once they have descended into it.
The difficulty of global optimization scales with the dimensionality of the search space and the complexity of the objective function landscape. For well-behaved convex functions, every local minimum is also the global minimum, and standard gradient descent reliably finds it. For non-convex functions with many local optima, global optimization methods must balance exploration of new regions against exploitation of promising regions already found. Evolutionary algorithms including genetic algorithms achieve this balance through population diversity. Bayesian optimization models the objective function with a probabilistic surrogate and uses the uncertainty in that model to guide exploration toward regions most likely to contain the global optimum. Simulated annealing probabilistically accepts worse solutions early in the search to escape local optima, gradually reducing this tolerance as the search progresses.
Many practical optimization problems in advertising and machine learning are not provably globally solved; instead, the goal is finding a solution good enough to be useful within a practical computational budget. The engineering judgment is how much computational investment is justified by the value of finding a better solution versus accepting a good local solution found quickly. For hyperparameter search on a client model, global optimization may be worth the compute cost if the performance improvement compounds over many inference runs. For a one-off media allocation, the value of marginal improvement must be weighed against the cost of running a more thorough search.
Every automated optimization system an agency uses, whether for budget allocation, bid management, creative testing, or model configuration, is solving an optimization problem. A working ad agency that understands whether those systems are finding genuinely good solutions or just locally optimal ones is better positioned to evaluate vendor claims, identify when an optimization system is underperforming relative to what the problem structure permits, and make informed tradeoffs between solution quality and computational cost.
Media allocation problems are almost always non-convex. The relationship between budget allocation across channels and campaign outcomes is not a smooth bowl-shaped surface with one clear bottom; it is a rugged landscape with diminishing returns, interaction effects between channels, and threshold behaviors. Algorithms that optimize allocation by gradient methods starting from the current allocation will find locally better allocations but miss globally better configurations that require reallocating in ways that temporarily look worse before they look better. Understanding this explains why allocation optimization tools with diverse search strategies consistently outperform those that only hill-climb from the current state.
Hyperparameter search quality directly determines model quality for clients. The performance of a custom propensity model or attribution model built for a client is bounded by how well the hyperparameter search found the global region of the configuration space. Grid search covers the space uniformly but wastes budget on obviously bad regions. Random search improves coverage but is still unguided. Bayesian optimization builds a model of which hyperparameter regions are promising and focuses search there, consistently finding better configurations than random or grid search on the same compute budget. Agencies that use Bayesian or evolutionary hyperparameter search build better models for the same time investment.
Platform bid optimization is a local search that can be globally suboptimal. Automated bidding systems in major ad platforms use gradient-based or rule-based methods that optimize from the current bid configuration. These systems can stabilize in locally optimal bid strategies that would improve substantially with a broader reset of the bid landscape. Periodic bid strategy audits that restart optimization from multiple starting points, rather than relying entirely on platform-side incremental optimization, sometimes reveal that the current stable strategy is far from the global best available within the platform’s optimizer.
An agency is building a custom lead scoring model for a SaaS client and using random search over 6 hyperparameters across a gradient boosted tree. After 100 iterations of random search, the best model achieves an AUC of 0.81 on the validation set. The agency switches to Bayesian optimization using a Gaussian process surrogate model that learns which regions of the hyperparameter space produce high validation AUC and focuses subsequent search iterations there. After an additional 60 Bayesian optimization iterations, the best configuration improves to an AUC of 0.87. The Bayesian search found a region of the hyperparameter space, specifically a combination of high learning rate, low subsampling, and deep trees, that random search had never sampled because the probability of landing in that region by chance across 100 random draws was low. The 0.06 AUC improvement translates to a 9% improvement in precision at the top lead decile, which the client’s sales team uses as their primary working list.
The automations and agents module covers how to build AI-powered optimization workflows that use global search methods for media allocation and model configuration, including when the computational cost of global search is justified by the value of the improvement.