Optimistic initial value analysis in a greedy selection approach to MAB problems

Abstract

The Multi Arm Bandit (MAB) problem is a well-known decision making problem, where the gambler (operator), seeks the highest value and best choice among arms with different reward distributions. In recent years, many effective strategies have been proposed for the mentioned MAB problem. The strategy's procedure mainly includes attempts on acquiring a better exploitation and efficiency, while also maximizing reliability by effectively exploring the search space. Hence, a suitable tradeoff between exploration and exploitation has been deemed necessary for achieving highest expected rewards. In the presented paper, the effect of assigning initial values has been studied as a deciding factor in the aforementioned tradeoff. In this regard, the initial value has been studied as an implicit exploration agent in greedy based selection strategies. Simulation and comparison of five methods have been presented and evaluated, using a similar exemplary case. Contrary to common belief, the obtained results have illustrated the high efficiency and exploration capability with proper initializing.