Analysis of (1+1) Evolutionary Algorithm and Randomized Local Search with Memory

• Published in: Evolutionary computation Vol. 19; no. 2; pp. 287 - 323
• Main Authors: Sung, Chi Wan, Yuen, Shiu Yin
• Format: Journal Article
• Language: English
• Published: One Rogers Street, Cambridge, MA 02142-1209, USA MIT Press 01.06.2011
• Subjects: (1+1) EA; Algorithms; Computer Simulation; Evolutionary algorithms; expected optimization time; Fitness; Invariants; Mathematical analysis; Mathematical models; Memory; Models, Genetic; Models, Theoretical; randomized local search; Search Engine; Searching; Tables (data); tabu search
• ISSN: 1063-6560; 1530-9304; 1530-9304
• DOI: 10.1162/EVCO_a_00029

This paper considers the scenario of the (1+1) evolutionary algorithm (EA) and randomized local search (RLS) with memory. Previously explored solutions are stored in memory until an improvement in fitness is obtained; then the stored information is discarded. This results in two new algorithms: (1+1) EA-m (with a raw list and hash table option) and RLS-m+ (and RLS-m if the function is a priori known to be unimodal). These two algorithms can be regarded as very simple forms of tabu search. Rigorous theoretical analysis of the expected time to find the globally optimal solutions for these algorithms is conducted for both unimodal and multimodal functions. A unified mathematical framework, involving the new concept of spatially invariant neighborhood, is proposed. Under this framework, both (1+1) EA with standard uniform mutation and RLS can be considered as particular instances and in the most general cases, all functions can be considered to be unimodal. Under this framework, it is found that for unimodal functions, the improvement by memory assistance is always positive but at most by one half. For multimodal functions, the improvement is significant; for functions with gaps and another hard function, the order of growth is reduced; for at least one example function, the order can change from exponential to polynomial. Empirical results, with a reasonable fitness evaluation time assumption, verify that (1+1) EA-m and RLS-m+ are superior to their conventional counterparts. Both new algorithms are promising for use in a memetic algorithm. In particular, RLS-m+ makes the previously impractical RLS practical, and surprisingly, does not require any extra memory in actual implementation.