Accelerated -Greedy Multi Armed Bandit Algorithm for Online Sequential-Selections Applications

Abstract

Current algorithms for solving multi-armed bandit (MAB) problem in stationary observations often perform well. Although this performance may be acceptable with accurate parameter settings, most of them degrade under non stationary observations. We setup an incremental e-greedy model with stochastic mean equation as its action-value function which is more applicable to real-world problems. Unlike the iterative algorithms suffering from step size dependency, we propose an adaptive step-size model (ASM) to introduce adaptive MAB algorithm. The proposed model employs e-greedy approach as action selection policy. In addition, a dynamic exploration parameter e is introduced to be ineffective by increasing decision maker's intelligence. The proposed model is empirically evaluated and compared with existing algorithms including the standard e-greedy, Softmax, e-decreasing and UCB-Tuned models under stationary as well as non stationary situations. ASM not only addresses concerns in parameter dependency problem but also performs either comparable or better than mentioned algorithms. Applying these enhancements to the standard e-greedy reduce the learning time which is more attractive to the wide range of on-line sequential selection-based applications such as autonomous agents, adaptive control, industrial robots and forecasting trend problems in management and economics domains.