Publication

Taeyoung Yun, Kiyoung Om, Jaewoo Lee, Sujin Yun, Jinkyoo Park

Abstract

Optimizing high-dimensional and complex blackbox functions is crucial in numerous scientific applications. While Bayesian optimization (BO) is a powerful method for sample-efficient optimization, it struggles with the curse of dimensionality and scaling to thousands of evaluations. Recently, leveraging generative models to solve black-box optimization problems has emerged as a promising framework. However, those methods often underperform compared to BO methods due to limited expressivity and difficulty of uncertainty estimation in high-dimensional spaces. To overcome these issues, we introduce DiBO, a novel framework for solving high-dimensional blackbox optimization problems. Our method iterates two stages. First, we train a diffusion model to capture the data distribution and deep ensembles to predict function values with uncertainty quantification. Second, we cast the candidate selection as a posterior inference problem to balance exploration and exploitation in high-dimensional spaces. Concretely, we fine-tune diffusion models to amortize posterior inference. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines across synthetic and realworld tasks. Our code is publicly available here.

#Diffusion Model #GFlowNet #Blackbox Optimization

자세히보기

Taeyoung Yun, Sujin Yun, Jaewoo Lee, Jinkyoo Park

Abstract

Optimizing complex and high-dimensional black-box functions is ubiquitous in science and engineering fields. Unfortunately, the online evaluation of these functions is restricted due to time and safety constraints in most cases. In offline model-based optimization (MBO), we aim to find a design that maximizes the target function using only a pre-existing offline dataset. While prior methods consider forward or inverse approaches to address the problem, these approaches are limited by conservatism and the difficulty of learning highly multi-modal mappings. Recently, there has been an emerging paradigm of learning to improve solutions with synthetic trajectories constructed from the offline dataset. In this paper, we introduce a novel conditional generative modeling approach to produce trajectories toward high-scoring regions. First, we construct synthetic trajectories toward high-scoring regions using the dataset while injecting locality bias for consistent improvement directions. Then, we train a conditional diffusion model to generate trajectories conditioned on their scores. Lastly, we sample multiple trajectories from the trained model with guidance to explore high-scoring regions beyond the dataset and select high-fidelity designs among generated trajectories with the proxy function. Extensive experiment results demonstrate that our method outperforms competitive baselines on Design-Bench and its practical variants.

#Diffusion Model #Blackbox Optimization

자세히보기

Jaewoo Lee, Sujin Yun, Taeyoung Yun, Jinkyoo Park

Abstract

Offline Reinforcement Learning (Offline RL) presents challenges of learning effective decision-making policies from static datasets without any online interactions. Data augmentation techniques, such as noise injection and data synthesizing, aim to improve Q-function approximation by smoothing the learned state-action region. However, these methods often fall short of directly improving the quality of offline datasets, leading to suboptimal results. In response, we introduce GTA, Generative Trajectory Augmentation, a novel generative data augmentation approach designed to enrich offline data by augmenting trajectories to be both high-rewarding and dynamically plausible. GTA applies a diffusion model within the data augmentation framework. GTA partially noises original trajectories and then denoises them with classifier-free guidance via conditioning on amplified return value. Our results show that GTA, as a general data augmentation strategy, enhances the performance of widely used offline RL algorithms across various tasks with unique challenges. Furthermore, we conduct a quality analysis of data augmented by GTA and demonstrate that GTA improves the quality of the data.

#Diffusion Model #Offline Reinforcement Learning

자세히보기

Soo Kyung Kim, Chihyeon Song, Weizhe Chen, Jinkyoo Park, Saman Mostafavi

Abstract

We propose a learning framework for interpretable HVAC control in buildings using deep reinforcement learning (DRL). Our framework includes a data-driven surrogate environment to emulate building dynamics and a Deep Symbolic Policy for discovering interpretable control policies. We focus on maintaining the temperature within the desired range for occupant comfort.
Our results show that the discovered symbolic policies are interpretable and perform well compared to standard DRL algorithms. Additionally, the discovered policies in surrogate models
exhibit transferability to physics-based environments with minimal performance degradation.

#Reinforcement Learning #HVAC

자세히보기

Saman Mostafavi, Chihyeon Song, Aayushman Sharma, Raman Goyal, Alejandro Brito

Abstract

We present a data-driven modeling and control framework for physics-based building emulators. Our approach consists of: (a) Offline training of differentiable surrogate models that accelerate model evaluations, provide costeffective gradients, and maintain good predictive accuracy for the receding horizon in Model Predictive Control (MPC), and (b) Formulating and solving nonlinear building HVAC MPC problems. We extensively evaluate the modeling and control performance using multiple surrogate models and optimization frameworks across various
test cases available in the Building Optimization Testing Framework (BOPTEST). Our framework is compatible with other modeling techniques and can be customized with different control formulations, making it adaptable and future-proof for test cases currently under development for BOPTEST. This modularity provides a path towards prototyping predictive controllers in large buildings, ensuring scalability and robustness in real-world applications.

#Model Predictive Control #HVAC

자세히보기

Kundo Park, Chihyeon Song,  Jinkyoo Park, Seunghwa Ryu

Abstract

The hierarchical structures found in biological materials lead to an outstanding balance of multiple material properties, and numerous research studies have been initiated to emulate the key concepts for the designing of engineering materials, the so-called bioinspired composites. However, the optimization of bioinspired composites has long been difficult as it usually falls into the category of ‘black-box problem’, the objective functions not being available in a functional form. Also, bioinspired composites possess multiple material properties that are ina trade-off relationship, making it impossible to reach a unique optimal design solution. As a breakthrough, we propose a data-driven material design framework which can generate bioinspired composite designs with an optimal balance of material properties. In this study, a nacreinspired composite is chosen as the subject of study and the optimization framework is applied to determine the designs that have an optimal balance of strength, toughness, and specific volume. Gaussian process regression was adopted for the modeling of a complex input–output relationship, and the model was trained with the data generated from the crack phase-field simulation. Then, multi-objective Bayesian optimization was carried out to determine pareto-optimal composite designs. As a result, the proposed data-driven algorithm could generate a 3D pareto surface of optimal composite design solutions, from which a user can choose a design that suits his/her requirement. To validate the result, several pareto-optimal designs are built using a PolyJet 3D printer, and their tensile test results show that each of the characteristic
designs is well optimized for its specific target objective.

#Bayesian Optimization #Design Automation

자세히보기

Kundo Park, Youngsoo Kim, Minki Kim, Chihyeon Song, Jinkyoo Park, Seunghwa Ryu

Abstract

The staggered platelet composite structure, one of the most well-known examples of biomimetics, is inspired by the microstructure of nacre, where stiff mineral platelets are stacked with a small fraction of soft polymer in a brick-and-mortar style. Significant efforts have been made to establish a framework for designing a staggered
platelet pattern that achieves an excellent balance of toughness and stiffness. However, because no analytical formula for accurately predicting its toughness is available because of the complexity of the failure mechanism of realistic composites, existing studies have investigated either idealized composites with simplified material properties or realistic composites designed by heuristics. In the present study, we propose a Bayesian optimization framework to design a staggered platelet structure that renders high toughness. Gaussian process regression (GPR)
was adopted to model statistically the complex relationship between the shape of the staggered platelet array and the resultant toughness. The Markov chain Monte Carlo algorithm was used to determine the optimal kernel hyperparameter set for the GPR. Starting with 14 initial training data collected with uniaxial tensile tests, a GPRbased Bayesian optimization using the expected improvement (EI) acquisition function was carried out. As a result, it was possible to design a staggered platelet pattern with a toughness 11% higher than that of the best
sample in the initial training set, and this improvement was achieved after only three iterations of our optimization cycle. As this optimization framework does not require any material theories and models, this process can be easily adapted and applied to various other material optimization problems based on a limited set of experiments
or computational simulations.

#Bayesian Optimization #Design Automation

자세히보기