Many-Objective Visual Analytics: Rethinking the Design of Complex Engineered Systems

Abstract

Over the past decade our research group has worked to operationlize our “many-objective visual analytics” (MOVA) framework for the design and management of complex engineered systems. Successful applications include urban water portfolio planning, satellite constellation design, airline scheduling, and product family design. The MOVA framework has four core components: (1) elicited problem conception and formulation, (2) many-objective search, (3) interactive visualization, and (4) negotiated design selection. Problem conception and formulation is the process of abstracting a practical design problem into a mathematical representation. We build on the emerging work in visual analytics to exploit interactive visualization of both the design space and the objective space in multiple heterogeneous linked views that permit exploration and discovery. Many-objective search produces a Pareto-approximate set of solutions from problem formulations that consider up to ten objectives based on current computational search capabilities. Negotiated design selection uses interactive visualization, reformulation, and optimization to discover desirable designs for implementation. Each of the activities in the framework is subject to feedback, both within the activity itself and from the other activities in the framework. These feedback processes transition formerly marginalized activities of reformulating the problem, refining the conceptual model of the problem, and refining the optimization, to represent the most critical process for innovating real world systems (i.e., learning how to frame the problems themselves). This study demonstrates insights gained by evolving the formulation of a General Aviation Aircraft (GAA) product family design problem. This problem’s considerable complexity and difficulty, along with a history encompassing several formulations, make it well-suited to demonstrate the MOVA framework. Our MOVA framework results compare a single objective, a two objective, and a ten objective formulation for optimizing the GAA product family. Highly interactive visual analytics are exploited to demonstrate how decision biases can arise for lower dimensional, highly aggregated problem formulations. As part of our efforts to operationlize the MOVA framework, we have also created rigorous search diagnostics to distinguish the efficiency, controllability, reliability, and effectiveness of multiobjective evolutionary algorithms (MOEAs). These diagnostics have distinguished the auto-adaptive behavior of our recently introduced Borg MOEA relative to a broad sampling of traditional MOEAs when addressing the GAA product family design problem.