Local Optima

A local optimum is a state in a system's fitness landscape that is superior to all immediately neighboring states but inferior to some distant state. Local optima are the trapdoors of evolution and optimization: they attract systems toward stability, then prevent them from reaching better stability. The problem is structural: to escape a local optimum, a system must pass through a region of lower fitness, and selection pressure — which rewards immediate improvement — systematically punishes such passages.

Local optima appear in biological evolution (body plans that work well enough to persist but block exploration of better designs), in technology (the QWERTY keyboard, locked in by path dependence), and in machine learning (neural networks that converge on suboptimal weight configurations). The severity of the local optima problem depends on the ruggedness of the fitness landscape: smooth landscapes have few optima and easy escapes; rugged landscapes have many optima and narrow valleys between them.

Escaping local optima requires mechanisms that temporarily suspend the pressure for immediate improvement: noise (random mutations, simulated annealing), diversity (populations that explore multiple peaks simultaneously), or exogenous shocks (environmental change that destabilizes the current optimum). Without such mechanisms, systems become frozen in states that are good enough to survive but not good enough to thrive.

The local optima problem is not a failure of search algorithms. It is a structural property of selection itself: selection is a hill-climbing mechanism, and hill-climbing cannot cross valleys. Any system that relies solely on selection — biological, economic, or computational — will eventually stagnate in a local optimum unless it has access to mechanisms that violate the logic of immediate improvement.