Abstract :

Mechanical Components Like Springs, Brackets, Beams And Turbine Blades Are Almost Never Designed For One Purpose. A Lighter Part Is Not Always The Safest Part, And The Most Cost Effective To Produce Is Not Always The Strongest Part, And So The Practicing Engineer Is Left With A Balancing Act Between Weight, Stress, Deflection, And Cost, And With A Set Of Structural Requirements That Cannot Be Broken. This Trade-off Is Reasonably Able To Be Managed By Classical Numerical Optimizers And Evolutionary Algorithms, But Requires Re-running The Algorithm Each Time The Load Case, Material Or Constraint Set Is Altered, And Each New Design Instance Is Treated As A Problem To Be Solved In Isolation. This Paper Considers An Alternative: A Design Agent That Learns A Reusable Constraint-aware Policy Over The Design Variables, Not Just One Specific Point, And That Can Be Conditioned On A Scalarisation Weight To Trace An Entire Trade Off Curve Instead Of One Design At A Time – Deep Reinforcement Learning (DRL). We Survey 40 Recent Works (2019-2026) Employing DRL In Multi-objective And Constrained Structural Design, Classify The Works By Contribution, And Choose To Use This Classification To Motivate A Constraint-first Formulation Of The Reward Which Is Combined With A Policy-gradient/evolution-strategy Training Loop And A Constrained-domination Pareto Archive. The Framework Is Instantiated And Tested On A Classical, Highly Constrained Coiling Spring Sizing Problem With Two Conflicting Objectives (weight And Shear Stress), And Has Six Structural Constraints, And The Results Are Compared With An NSGA-II Based Genetic Algorithm And Unguided Random Search In Five Independent Runs. The Lightweight DRL Policy Required About An Order Of Magnitude Fewer Design Evaluations Than Random Search To Converge To Its Own Hypervolume, And Was Able To Satisfy Structural Constraints About Sixty Times More Consistently Than Random Sampling, But A Well-tuned Evolutionary Baseline Policy Was Still Able To Find A Lower Weight Feasible Design Within The Same Number Of Design Evaluations. We Provide All Five Of The Comparison Metrics And Address Why Each One Is Better In Some Areas, As Well As What A Productionscale DRL Design Agent Would Require That We Did Not Include In This Proof-of-concept.

Published:

16-3-2021

Issue:

Vol. 21 No. 3 (2021)

Page Nos:

27 - 43

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite