AntNet with Reward-Penalty Reinforcement Learning

Abstract

The paper deals with a modification in the learning phase of AntNet routing algorithm, which improves the system adaptability in the presence of undesirable events. Unlike most of the ACO algorithms which consider reward-inaction reinforcement learning, the proposed strategy considers both reward and penalty onto the action probabilities. As simulation results show, considering penalty in AntNet routing algorithm increases the exploration towards other possible and sometimes much optimal selections, which leads to a more adaptive strategy. The proposed algorithm also uses a self-monitoring solution called Occurrence-Detection, to sense traffic fluctuations and make decision about the level of undesirability of the current status. The proposed algorithm makes use of the two mentioned strategies to prepare a self-healing version of AntNet routing algorithm to face undesirable and unpredictable traffic conditions.