Design Strategy Network: A deep hierarchical framework to represent generative design strategies in complex action spaces

Abstract

Generative design problems often encompass complex action spaces that may be divergent over time, contain state-dependent constraints, or involve hybrid (discrete and continuous) domains. To address those challenges, this work introduces Design Strategy Network (DSN), a data-driven deep hierarchical framework that can learn strategies over these arbitrary complex action spaces. The hierarchical architecture decomposes every action decision into first predicting a preferred spatial region in the design space and then outputting a probability distribution over a set of possible actions from that region. This framework comprises a convolutional encoder to work with image-based design state representations, a multi-layer perceptron to predict a spatial region, and a weight-sharing network to generate a probability distribution over unordered set-based inputs of feasible actions. Applied to a truss design study, the framework learns to predict the actions of human designers in the study, capturing their truss generation strategies in the process. Results show that DSNs significantly outperform non-hierarchical methods of policy representation, demonstrating their superiority in complex action space problems.

Learning to Design Without Prior Data: Discovering Generalizable Design Strategies Using Deep Learning and Tree Search

“…They use a sampling-based action-independent formulation and enable a new policy iteration method. Our current research provides a much simpler approach that extends the WU-UCT formulation [38] by employing an order-invariant policy network [14] to sample feasible actions. 6…”

“…Section 3.2 details the deep learning part which represents generative design strategies in complex action spaces. This deep learning network is built upon Design Strategy Network (DSN) [14] and enables SLDA to learn over selfgenerated experience and extract meaningful relationships between the state and selected actions. However, in this work this framework is extended from a descriptive model to a generative model as it is used to predict actions on sequential design states.…”

“…This section introduces a unique complex action Markov Decision Process (MDP) that can define a generative design process as a sequential process [14,55,56]. This definition provides a standardized manner of representing design processes and guiding environment development for learning-based systems.…”

“…We propose a sampling-based approach to discretize the continuous domains and limit the feasible actions [14]. A prior probability 𝑝 ) is defined to sample in the continuous parameter domain 𝑎 ) "…”

“…The action space of a design process is challenging to generically represent as it constrains the problem space and dictates the exploration capability of a designer. Computationally, the action space may include a combination of continuous and discrete parameters; they may be constrained, state-dependent, or even be temporally divergent, as the number of actions keep on increasing as the process progresses [14]. An action definition that can represent all such variability is referred to as a complex action space [15,16].…”