In budget–limited multi–armed bandit (MAB) problems, thelearner’s actions are costly and constrained by a fixed budget.Consequently, an optimal exploitation policy may not be topull the optimal arm repeatedly, as is the case in other variantsof MAB, but rather to pull the sequence of different arms thatmaximises the agent’s total reward within the budget. Thisdifference from existing MABs means that new approachesto maximising the total reward are required. Given this, wedevelop two pulling policies, namely: (i) KUBE; and (ii)fractional KUBE. Whereas the former provides better performanceup to 40% in our experimental settings, the latteris computationally less expensive. We also prove logarithmicupper bounds for the regret of both policies, and show thatthese bounds are asymptotically optimal (i.e. they only differfrom the best possible regret by a constant factor).

We introduce the budget–limited multi–armed bandit (MAB), which captures situations where a learner’s actions are costly and constrained by a fixed budget that is incommensurable with the rewards earned from the bandit machine, and then describe a first algorithm for solving it. Since the learner has a budget, the problem’s duration is finite. Consequently an optimal exploitation policy is not to pull the optimal arm repeatedly, but to pull the combination of arms that maximises the agent’s total reward within the budget. As such, the rewards for all arms must be estimated, because any of them may appear in the optimal combination. This difference from existing MABs means that new approaches to maximising the total reward are required. To this end, we propose an epsilon–first algorithm, in which the first epsilon of the budget is used solely to learn the arms’ rewards (exploration), while the remaining 1 − epsilon is used to maximise the received reward based on those estimates (exploitation). We derive bounds on the algorithm’s loss for generic and uniform exploration methods, and compare its performance with traditional MAB algorithms under various distributions of rewards and costs, showing that it outperforms the others by up to 50%.