THE ALGORITHM PROCUREMENT PROBLEM IN ARTIFICIAL INTELLIGENCE

Abstract

Artificial intelligence systems can often solve the same task using substantially different computational methods. Depending on the problem, these may include rule-based and elimination procedures, lookup methods, direct algebraic solutions, pseudoinverse and regularized methods, statistical classifiers, and iterative optimization methods such as gradient descent (GD), stochastic gradient descent (SGD), conjugate gradient (CG), and Adam. The existence of alternative algorithms raises an engineering question that is distinct from whether a method can solve the task: how should an AI system select among several adequate computational methods? This paper introduces the Algorithm Procurement Problem (APP), in which alternative computational methods are treated as competing candidates for an AI task. APP does not prescribe a particular algorithm and does not favor iterative or non-iterative computation. Instead, it imposes an algorithm-selection procedure. Candidate methods are generated, their adequacy for the requested task is evaluated, inadequate candidates are rejected, and the computational prices of the remaining candidates are compared before the task is awarded. Computational price is defined by the application rather than by a universal measure. It may include setup and training computation, iterations, inference cost, memory and data movement, hardware utilization, energy, latency, reliability, numerical stability, repeated-use cost, and, where relevant, mission or operational requirements. Consequently, the least-cost method is not necessarily the method requiring the fewest operations or the least energy; it is the adequate method providing the best justified computational choice under the requirements of the task. The motivation for APP is examined through exploratory ordinary-user experiments in which AI systems are allowed to select their own computational methods without an algorithm being prescribed. The experiments indicate that computationally economical alternatives may already be available to the AI system but may become explicit only when computational price is introduced into the selection request. This observation motivates a distinction between knowing an algorithm and selecting it. APP therefore proposes that computational economy and operational requirements should participate in algorithm selection rather than being considered only after a computational paradigm has been selected. An iterative method may win when it provides the best justified solution, while a direct, lookup, rule-based, or elimination method may win when it adequately performs the same task at a lower justified computational price. The central proposition is that the algorithm should be treated as a bidder, while the selection procedure serves as the procurement authority.