Acceptance Driven Local Search and Evolutionary Algorithms

Local Search (LS) and Evolutionary Algorithms (EA) are probabilistic search algorithms, widely used in global optimization, where selection is important as it drives the search. In this paper, we introduce acceptance, a metric measuring the selective pressure in LS and EA, that is the tradeoff between exploration and exploitation. Informally, acceptance is the proportion of accepted non-improving transitions in a selection. We propose a new LS algorithm, SAad, based on acceptance schedule (a schedule for the selective pressure). In EA, two new selection rules based on the Metropolis criterion are introduced. They allow two new EA (2MT and RT) based on acceptance schedule. They demonstrate a possible way of merging LS and EA technologies. Benchmarks show that the developed algorithms are more performant than standard SA and EA algorithms, and that SAad is as efficient as the best SA algorithms while 2MT and RT are complementary to Evolution Strategies.