Fricke, Clemens David

Wolff, Daniel

Kemmerling, Marco

Elgeti, Stefanie

2024-01-04T10:42:16Z

2024-01-04T10:42:16Z

2023-06-21

<div class="csl-bib-body"> <div class="csl-entry">Fricke, C. D., Wolff, D., Kemmerling, M., &#38; Elgeti, S. (2023, June 21). <i>Investigation of Reinforcement Learning in Shape Optimization</i> [Conference Presentation]. 11th International Conference on IsoGeometric Analysis (IGA 2023), Lyon, France. http://hdl.handle.net/20.500.12708/191068</div> </div>

http://hdl.handle.net/20.500.12708/191068

Reinforcement Learning (RL) is a subfield of Machine Learning that aims to mimic the way how humans learn to accomplish a new task. This process is mostly guided by experience, i.e., by trying out how to act in a specific situation in order to achieve a certain goal. Once enough experience has been gathered, it can be utilized to make more educated decisions to accomplish this task efficiently. In RL terminology, this problem can be stated as follows: An agent is assigned the task to solve a given problem in multiple sequential steps. It does this by observing states from the environment and generating actions in response to these observations. The sequence of steps required to solve the problem is called an episode. The selection of actions in each step is defined by a policy, i.e., a mapping from observations to actions, which is learned through interactions with the environment such that the cumulative reward received for the actions is maximized. The agent can hence be considered as a learning algorithm that optimizes a policy for solving the respective task, which is encoded by its environment. In this work [1], we apply RL as a shape optimization driver. Based on a variety of basic examples lent from production engineering, we were able to show that an RL agent can indeed learn to optimize a geometry. In particular, we compare a variety of algorithms, considering both the direct optimization approach and the incremental optimization approach [2]. These differ in how the agent utilizes its actions to modify the degrees of freedom (DOFs) of the geometry parameterization. With incremental shape optimization, only a single DOF is changed per step. In the implementation used in our RL-based shape optimization package releso, the DOFs is either incremented or decremented by a predefined amount. In contrast, in the direct optimization approach, all DOF are adapted at once in every step, and can, at least in our implementation, be set to any value within a specified interval. This results in the following behaviour: A fully trained agent, following the incremental method, iteratively approaches the optimal geometry in multiple small steps. Contrasting, an agent following the direct strategy tries to find the optimal geometry in a single step. We are able to show that RL-algorithms are very suitable for repeated optimization tasks of similar structure. Imagine for example the wings for different types of airplanes.

en

Reinforcement Learning

shape optimization

Investigation of Reinforcement Learning in Shape Optimization

Presentation

Vortrag

RWTH Aachen University, Germany

RWTH Aachen University, Germany

Conference Presentation

C4

C6

C3

Mathematical and Algorithmic Foundations

Modeling and Simulation

Computational System Design

20

30

50

E317-01-2 - Forschungsgruppe Struktur- und Werkstoffsimulation

E317-01-1 - Forschungsgruppe Numerische Analyse- und Designmethoden

0000-0002-9495-2285

0000-0003-0141-2050

0000-0002-4474-1666

11th International Conference on IsoGeometric Analysis (IGA 2023)

18-06-2023

21-06-2023

On Site

Event for scientific audience

Lyon

FR

Elgeti, Stefanie

Maschinenbau

Informatik

Mathematik

2030

1020

1010

30

40

30

http://purl.org/coar/resource_type/c_18cp

conference paper not in proceedings

no Fulltext

en

none

Publications

E317-01-2 - Forschungsgruppe Struktur- und Werkstoffsimulation

E317-01 - Forschungsbereich Leichtbau

RWTH Aachen University

E317-01 - Forschungsbereich Leichtbau

0000-0002-9495-2285

0000-0003-0141-2050

0000-0002-4474-1666

E317-01 - Forschungsbereich Leichtbau

E317 - Institut für Leichtbau und Struktur-Biomechanik

E317 - Institut für Leichtbau und Struktur-Biomechanik