Effective Multi-caste Ant Colony System for Large Dynamic Traveling Salesperson Problems Leonor Melo 1,2 , Francisco Pereira 1,2 , and Ernesto Costa 2 1 Instituto Polit´ecnico de Coimbra, ISEC, DEIS, Rua Pedro Nunes, Quinta da Nora, 3030-199 Coimbra, Portugal 2 Centro de Inform´atica e Sistemas da Univ. Coimbra, 3030-790 Coimbra, Portugal leonor@isec.pt, xico@dei.uc.pt, ernesto@dei.uc.pt Abstract. Multi-caste ant algorithms allow the coexistence of different search strategies, thereby enhancing search effectiveness in dynamic op- timization situation. We present two new variants for a multi-caste ant colony system that promote a better migration of ants between alterna- tive behaviors. Results obtained with large and highly dynamic traveling salesperson instances confirm the effectiveness and robustness of the ap- proach. A detailed analysis reveals that one of the castes should adopt a clearly exploratory behavior, as this minimizes the recovery time after an environmental change. Keywords: Ant Colony Optimization, Dynamic Traveling Salesperson Problem, Multi-caste Ant Colony System, Traffic factor 1 Introduction Ant Colony Optimization (ACO) encompasses a class of algorithms loosely in- spired in the behavior of ants [4]. First developed to deal with the Traveling Salesperson Problem, it has proven successful in a wide range of hard combina- torial optimization problems [4]. Ant Colony System (ACS) [3] is one of the most successful ACO variants and its main distinguishing feature is the existence of a greedy decision rule adopted by artificial ants when building a solution for the problem being solved. ACS, as well as other ACO variants, depend on a set of parameters that govern the way the search is conducted. Although beneficial, a careful adjustment of the settings is far from trivial. Also, the ideal setting may change throughout an optimization run, as the search conditions vary. In dynamic environments, where the problem modifies over time, this situation is amplified. In a previous work [15] we proposed a multi-caste framework that allows the coexistence of different sets of parameter values, hence search strategies, inside a single ACS algorithm. Also, although the total colony size is fixed, ants may migrate between castes during the run, thereby favoring the specific search strategy that seems to be more suitable at a given period. In [16], the multi-caste ACS was applied to several Dynamic Traveling Salesperson Problem (DTSP) instances. Results revealed that the adoption of different castes enhances the robustness of the