DECISION MAKING IN DYNAMIC ENVIRONMENTS AN APPLICATION OF MACHINE LEARNING TO THE ANALYTICAL HIERARCHY PROCESS

Abstract

The purpose of this work is to propose a method of algorithmic decision making that builds on the Analytical Hierarchy Process by applying reinforcement learning. Decision making in dynamic environments requires adaptability as new information becomes available. The Analytical Hierarchy Process (AHP) provides a method for comparative decision making but is insufficient to handle information that becomes available over time. Using the opinions of one or many subject matter experts and the AHP, the relative importance of evidence can be quantified. However, the ability to explicitly measure the interdependencies is more challenging. The interdependency between the different evidence can be exploited to improve the model accuracy, particularly when information is missing or uncertain. To establish this ability within a decision-making tool, the AHP method can be optimized through a stochastic gradient descent algorithm. To illustrate the effectiveness of the proposed method, an experiment was conducted on air target threat classification in time series developing scenarios.