We propose a new distance algorithm for phylogenetic estimation based on Ant Colony Optimization (ACO), named Ant-Based Phylogenetic Reconstruction (ABPR). ABPR joins two taxa iteratively based on evolutionary distance among sequences, while also accounting for the quality of the phylogenetic tree built according to the total length of the tree. Similar to optimization algorithms for phylogenetic estimation, the algorithm allows exploration of a larger set of nearly optimal solutions. We applied the algorithm to four empirical data sets of mitochondrial DNA ranging from 12 to 186 sequences, and from 898 to 16,608 base pairs, and covering taxonomic levels from populations to orders. We show that ABPR performs better than the commonly used Neighbor-Joining algorithm, except when sequences are too closely related (e.g., population-level sequences). The phylogenetic relationships recovered at and above species level by ABPR agree with conventional views. However, like other algorithms of phylogenetic estimation, the proposed algorithm failed to recover expected relationships when distances are too similar or when rates of evolution are very variable, leading to the problem of long-branch attraction. ABPR, as well as other ACO-based algorithms, is emerging as a fast and accurate alternative method of phylogenetic estimation for large data sets.

Este trabalho apresenta um algoritmo de roteamento adaptativo, denominado Agentes-Q, baseado na integração de três estratégias de aprendizagem combinadas com alguns mecanismos para aumentar sua velocidade de adaptação. As três estratégias são: aprendizagem-Q, aprendizagem por reforço dual e aprendizagem baseada no comportamento de colônias de formigas. O algoritmo proposto é composto por um conjunto de agentes móveis simples, que se comunicam indiretamente e cooperam na busca dos melhores caminhos na rede. Os agentes selecionam as rotas de forma distribuída e atualizam incrementalmente as informações utilizadas na escolha dos caminhos. Agentes-Q foi aplicado a uma rede telefônica da British Telecom e a porcentagem média de chamadas perdidas por ele foi comparada a de dois algoritmos baseados no comportamento de colônias de formigas. Os experimentos envolveram variações nos padrões de tráfego, nível de carga e topologia da rede e a utilização de ruído na informação usada na seleção das rotas. Agentes-Q obteve melhor desempenho que seus competidores, apresentando maior capacidade de adaptação às situações consideradas.

This paper presents an adaptive routing algorithm, called Q-Agents, based on the integration of three learning strategies combined with some mechanisms to increase its speed of adaptation. The strategies were: Q-learning, dual reinforcement learning and learning based on ant colony behavior. The proposed algorithm is composed of a set of simple mobile agents that communicate indirectly between themselves and cooperate to find the best paths through the network. The agents select the routes in a distributed way and update information used in this task incrementally. Q-Agents has been applied to a telephone network from British Telecom and the mean percentage of lost calls by it was compared with the percentage of two algorithms based on ant colony behavior. The experiments comprised variations of the network traffic patterns, load level and topology and use of noise in information to select the routes. Q-Agents performed better than its competitors, presenting higher capacity of adaptation to the considered situations.