maxNoImprov

MaxNoImprov is a stopping criterion used in iterative optimization and heuristic search to terminate a run when no improvement in the objective value has been observed for a specified number of consecutive iterations. It focuses on stagnation rather than elapsed time, enabling more efficient use of computational resources.

Formally, during an optimization run, the current best objective value is tracked. A counter is incremented each iteration that does not produce an improvement over this best value and is reset to zero when an improvement occurs. If the counter reaches the predefined threshold maxNoImprov, the search stops. Variants may define “improvement” with strict or relaxed criteria, or allow resets under certain conditions such as minor improvements.

MaxNoImprov is widely used in local search, hill-climbing, simulated annealing, genetic algorithms, and hyperparameter tuning. It helps prevent wasted effort in regions of the search space that yield no progress and provides a simple, problem-agnostic termination rule. The choice of the threshold is problem dependent: too small a value risks premature termination, while too large a value may waste time on diminishing returns. In practice, practitioners may combine maxNoImprov with other criteria, such as time limits, maximum iterations, or restart strategies, and may adapt the threshold dynamically based on observed progress or noise levels.

Related concepts include early stopping, patience parameters in machine learning, and stagnation limits. MaxNoImprov emphasizes robustness to minor fluctuations and non-monotonic objective landscapes.