Pengyuan Zhai

Keywords: Generative Adversarial Networks, Path Integral Monte Carlo, Stochastic Processes, Discrete Optimization, Queuing Networks, Enterprise-scale Simulations, Urban Systems Optimization, Structural Optimization, Structural Health Monitoring

---

Current Progress

Update as of Feb 12, 2021:

• New paper title, “Information Maximization and Contrastive Learning Generative Adversarial Network for Simulated Quantum Annealing” contains a revised model to replace the original design (old paper link). In the new model, I combined and tested various mutual information (lower bound) maximization strategies to the GAN framework to improve the generator’s data output quality at continuous target labels.
  
  Additionally, rigorous proofs are provided to give a unifying view on the following mutual information lower bound maximization methods: 1. variational lower bound through difference of entropies (DoE), 2. noice-contrastive estimation (NCE) methods, 3. Wasserstein Dependency Measure (WDM), by formulating them as an optimization problem involving maximum likelihood estimation on positive and/or negative samples.

• Final paper for French History Class, Hist 103B, titled “What Makes Satires Satirical? Discover Patterns from French Revolution Media with Deep Learning“. I used GAN to extract salient visual features from French Revolution era artworks (engravings, pamphlets, etc) and argued that political engravings use wide and complex scenes to showcase political tension, meanwhile pamphlet pictures use simple metaphorical motifs to comment on political events and public figures.

---

Papers and Manuscripts

• Balanced Batch Semi-supervised GAN Publication (BSS-GAN): Gao, Zhai and Mosalam (2020). Balanced Semi-supervised Generative Adversarial Network in Vision-based Structural Damage Assessment under the Low-data and Imbalanced-class Regime. In proceedings of the 17th World Conference on Earthquake Engineering (Conference Paper) and submitted to Computer-Aided Civil and Infrastructure Engineering (Journal Paper)
  Links: Journal Paper | Conference Paper | Code
  
  
• Text Analytics for Resilience GAN with Knowledge Distillation Publication (TAR-GAN): Tsai, Gunay, Li, Hwang, Zhai, El Ghaoui, Mosalam (2020). Text Analytics for Resilience-Enabled Extreme Events Reconnaissance. Submitted to NeurIPS 2020 Artificial Intelligence for Humanitarian Assistance and Disaster Response Workshop. Link: Paper | Code
  
  
• Simulate Path-Integral Markov Chain Monte Carlo Quantum Ising Distribution with GAN (SQA-GAN): Zhai (2020). Simulate Quantum Annealing Ising Model with Conditional Generative Adversarial Network. Link: Working Draft | Code
  
  
• Extract Semantic Features from French Revolution Era Media with GAN: Zhai (2020), What Makes Satires Satirical? Discover Patterns from French Revolution Media with Deep Learning. Link: Manuscript
  
  
• ECM Optimization Algorithm: Find the optimal combination of Energy Conservation Measure (ECM) packages for a building. Github Links: Summary | Code (Ruby with GLPK engine) | Code (Python with Gurobi engine)
  
  
• Linear Programming in Plastic Structural Analysis: Establish a physical connection between solving for maximum loads in a plastic system and the simplex algorithm in linear programming through a 3-element plane truss system. Manuscript

Structural Software

• Structural Engineering Software (FEDEAS): Finite Elements for Design, Evaluation and Analysis of Structures Web Solver (FEDEASLab Web). Link: app.fedeas.com

Urban System Geospatial Dataset Development

• R-tree Geo-data Fast Join: Fast join data records by matching geo-spatial multipolygons with the R-tree data structure and develop building energy benchmark datasets for Lawrence Berkeley National Laboratory’s City Building Energy Saver (CityBES). Links: Summary | Code | LBNL CityBES Software

Textbook Writing

• Cyber-Physical Modeling and Machine Learning Towards Smart Electrical Equipment Systems: Co-authored Part IV in Prof. Khalid Mosalam’s textbook on Data-driven Vision-based Structural Health Monitoring (page 19-23). Manuscript