Design as a sequential decision process with applications in structural engineering: Semm Seminar

Seminar | November 13 | 12-1 p.m. | 502 Davis Hall

 Gordon Warn, Associate Professor, Department of Civil & Environmental Engineering, The Pennsylvania State University, University Park

 Civil and Environmental Engineering (CEE)

The engineering design community is being tasked with generating designs that must satisfy ever more criteria, such as purchase cost, various performance-based metrics, life-cycle costs, among others, necessitating higher fidelity (and more computationally demanding) models to gain insight to resolve tradeoffs and find satisfing design alternatives. For such complex systems, a diverse set of design solutions exists that the designer must broadly explore to select the ideal design alternative.
This seminar presents an emerging design concept that closely couples set-based design with model-based simulation treating the design process formally as a sequential decision process (SDP). In this paradigm, mathematical models of increasing fidelity, Mj, are used in a sequence to successively provide tighter bounds on the decision criteria, facilitating the systematic contraction of the set of design alternatives; Xi, through a sequence of discrete decision states, until a choice set is obtained from which a design can be selected. The SDP using models to bound objectives is introduced, formally defined, and applied to the system-level design of seismic-resisting structural frames with deterministic decision criteria using the capacity spectrum method. In spite of its desirable features, the SDP using models to bound objectives has an important limitation. The multiple model fidelities and multiple discrete decision states result in a multitude of model sequences to arrive at a choice set, some sequences requiring significantly fewer model evaluations than others. It is desirable to select the most efficient sequence possible. However, the optimal sequence of model fidelities can only be determined, using for example dynamic programming, after constructing a complete dataset of the computational cost of executing each model fidelity on each design alternative and the associated bounds on the decision criteria, thus requiring an excessive number of model evaluations and limiting the practical application of the SDP. Recent efforts to overcome this limitation will be presented, whereby the SPD is reformulated as a finite Markov Decision Process that is solved using reinforcement learning (RL) to identify approximately optimal sequences of model fidelities using actual experienced transitions in place of knowledge of expected transitions.