Machine Decision Makers as a Laboratory for Interactive EMO

A key challenge, perhaps the central challenge, of multi-objective optimization is how to deal with candidate solutions that are ultimately evaluated by the hidden or unknown preferences of a human decision maker (DM) who understands and cares about the optimization problem. Alternative ways of addressing this challenge exist but perhaps the favoured one currently is the interactive approach (proposed in various forms). Here, an evolutionary multi-objective optimization algorithm (EMOA) is controlled by a series of interactions with the DM so that preferences can be elicited and the direction of search controlled. MCDM has a key role to play in designing and evaluating these approaches, particularly in testing them with real DMs, but so far quantitative assessment of interactive EMOAs has been limited. In this paper, we propose a conceptual framework for this problem of quantitative assessment, based on the definition of machine decision makers (machine DMs), made somewhat realistic by the incorporation of various non-idealities. The machine DM proposed here draws from earlier models of DM biases and inconsistencies in the MCDM literature. As a practical illustration of our approach, we use the proposed machine DM to study the performance of an interactive EMOA, and discuss how this framework could help in the evaluation and development of better interactive EMOAs.