Towards an Improved Strategy for Solving Multi Armed Bandit Problem
Semiu A. Akanmu1, Rakhen Garg2, Abdul Rehman Gilal3

1Semiu A. Akanmu Department of Computer Science North Dakota State University Fargo, USA.
2Rakhen Garg Department of Computer Science,North Dakota State University, Fargo, USA. E-mail:
3Abdul Rehman Gilal Department of Computer Science Sukkur IBA University, Sindh, Pakistan.

Manuscript received on September 13, 2019. | Revised Manuscript received on 24 September, 2019. | Manuscript published on October 10, 2019. | PP: 5060-5064 | Volume-8 Issue-12, October 2019. | Retrieval Number: L25221081219/2019©BEIESP | DOI: 10.35940/ijitee.L2522.1081219
Open Access | Ethics and Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Multi-Armed Bandit (MAB) problem is one of the classical reinforcements learning problems that describe the friction between the agent’s exploration and exploitation. This study explores metaheuristics as optimization strategies to support Epsilon greedy in achieving an improved reward maximization strategy in MAB. In view of this, Annealing Epsilon greedy is adapted and PSO Epsilon greedy strategy is newly introduced. These two metaheuristics-based MAB strategies are implemented with input parameters, such as number of slot machines, number of iterations, and epsilon values, to investigate the maximized rewards under different conditions. This study found that rewards maximized increase as the number of iterations increase, except in PSO Epsilon Greedy where there is a non-linear behavior. Our AnnealingEpsilon greedy strategy performed better than Epsilon Greedy when the number of slot machines is 10, but Epsilon greedy did better when the number of slot machines is 5. At the optimal value of Epsilon, which we found at 0.06, Annealing Epsilon greedy performed better than Epsilon greedy when the number of iterations is 1000. But at number of iterations ≥ 1000, Epsilon greedy performed better than Annealing Epsilon greedy. A stable reward maximization values are observed for Epsilon greedy strategy within Epsilon values 0.02 and 0.1, and a drastic decline at epsilon > 0.1.
Keywords: Multi-Armed Bandit Strategy, Reinforcement Learning, Metaheuristics, Epsilon Greedy, Annealing, Particle Swarm Optimization
Scope of the Article: Deep Learning