On th er elatio nb etween Ant Colony Optimization and Heuristically Accelerated Reinforcement Learning
Reinaldo A. C. Bianchi, Carlos H. C. Ribeiro, Anna Helena Reali Costa · 2009
This paper has two mai ng oals :t he firs ti s t op ropos ea new class of Heuristically AcceleratedReinforcement Learning algorithms (HARL), the Distributed HARLs, describing on ea lgorithm of this class, the Heuristically Accelerated Distributed QLearning (HADQL); and the second is to show that Ant Colony Optimization(ACO) algorithms ca nb e seenasinstancesofDistributedHARLsalgorithms. I np articular, this paper shows tha tt he Ant Colony System (ACS) algorithm ca nb e interpreted as a particular case of the HADQL algorithm. This interpretation is very attractive, as many of th ec onclusions obtained for RL algorithms remai nv alid for Distributed HARL algorithms,such as the guarantee of convergence to equilibrium. I no rder t ob etter evaluate the proposal, w ec ompared the performances of the Distributed Q-Learning, the HADQL and the ACS algorithms in the Traveling Salesman Problem domain. The result ss how that HADQL and the ACS algorithm have similar performances, as it woul db ee xpected from the hypothesis tha tt hey are, in fact, instances of the same class of algorithms.