Budget allocation using weakly coupled, constrained Markov decision processes
Craig E. Boutilier, Tyler Lu · Uncertainty in Artificial Intelligence · 2016
We consider the problem of budget (or other resource) allocation in sequential decision problems involving a large number of concurrently running sub-processes, whose only interaction is through their consumption of budget. Our first contribution is the introduction of budgeted MDPs (BMDPs), an MDP model in which policies/values are a function of available budget, (c.f. constrained MDPs which are solved given a fixed budget). BMDPs allow one to explicitly trade off allocated budget and expected value. We show that optimal value functions are concave, non-decreasing in budget, and piecewise-linear in the finite horizon case, and can be computed by dynamic programming (and support ready approximation). Our second contribution is a method that exploits BMDP solutions to allocate budget to a large number of independent BMDPs, coupled only by their common budget pool. The problem can be cast as a multiple-choice knapsack problem, which admits an efficient, optimal greedy algorithm. Empirical results in an online advertising domain confirm the efficacy of our methods.