Ant Q a reinforcement learning approach to the traveling salesman problem英文文献.pdfVIP

Appeared in: Proceedings of ML-95, Twelfth Intern. Conf. on Machine Learning, Morgan Kaufmann, 1995, 252–260. Ant-Q: A Reinforcement Learning approach to the traveling salesman problem Luca M. Gambardella Marco Dorigo IDSIA IRIDIA, Université Libre de Bruxelles Corso Elvezia 36 Avenue Franklin Roosevelt 50 6900 Lugano CP 194/6 Switzerland 1050 Bruxelles, Belgium, EU luca@idsia.ch mdorigo@ulb.ac.be http://www.idsia.ch http://iridia.ulb.ac.be/dorigo/dorigo.html Abstract Q family of algorithms is presented in Section 2, and a few different members of the family are discussed and In this paper we introduce Ant-Q, a family of experimentally compared in Section 3. Experiments are algorithms which present many similarities run on symmetric and asymmetric traveling salesman with Q-learning (Watkins, 1989), and which we problems. In Section 4 we experimentally study some apply to the solution of symmetric and asym- properties of Ant-Q. In particular we show that, as com- metric instances of the traveling salesman prob- putation goes on, agents do not converge toward a com