Physics-Driven ML-Based Modelling for Correcting Inverse Estimation
Authors: ruiyuan kang, Tingting Mu, Panagiotis Liatsis, Dimitrios Kyritsis
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | GEESE is tested on three real-world SAE inverse problems and compared to a number of state-of-the-art optimization/search approaches. Results show that it fails the least number of times in terms of finding a feasible state correction, and requires physical evaluations less frequently in general. |
| Researcher Affiliation | Collaboration | Ruiyuan Kang Bayanat AI Abu Dhabi, UAE ruiyuan.kang@bayanat.ai Tingting Mu University of Manchester Manchester, M13 9PL, United Kingdom tingting.mu@manchester.ac.uk Panos Liatsis Khalifa Univeristy Abu Dhabi, UAE panos.liatsis@ku.ac.ae Dimitrios C. Kyritsis Khalifa Univeristy Abu Dhabi, UAE dimitrios.kyritsis@ku.ac.ae |
| Pseudocode | Yes | Algorithm 1 Sketch of GEESE; Algorithm 2 GEESE |
| Open Source Code | Yes | Algorithm 2 outlines the pseudocode of GEESE2, while Fig.1 illustrates its system architecture. Our key implementation practice is summarized below. 2project repo: https://github.com/Ralph Kang/GEESE |
| Open Datasets | Yes | Following the same setting as in [42], the goal is to estimate the design parameters that can achieve the performance of a CFM-56 turbofan engine, for which the thrust should be 121 KN and the thrust specific fuel consumption should be 10.63 g/(k N.s) [67]."; "We have randomly selected 100 combinations of the safety factor and overall cost from the known Pareto front [58]"; "They correspond to 100 randomly selected combinations of the distortion and nonlinear factors from the known Pareto front in [59] |
| Dataset Splits | No | The paper mentions collecting initial training data D0 and continuously expanding it as Dt for training the surrogate error model. Early stopping is applied based on training loss ('training loss in Eq. (5) is smaller than a preidentified threshold ϵe'). However, it does not explicitly specify a separate validation dataset split. |
| Hardware Specification | No | The paper discusses 'computing time' and 'limited computing resource' but does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using specific packages like 'Bayesian Optimization' [11] and 'pymoo' [75] for implementing some of the compared methods, but it does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | The common hyperparameter settings shared between all three studied problems include Te = 40, ϵe = 1e 4 and N = 64, and the learning rates of 1e 2 and 1e 4, for training the exploitation generator and base neural networks, respectively. Different focus coefficients of c = 1.5, 2 and 5... an increasing training frequency coefficient δG = 1, 1 and 7 is used for problems 1, 2 and 3... The ensemble surrogate model for estimating the implicit errors is constructed as an average of 4 multilayer perceptrons (MLPs) each with three hidden layers consisting of 1024, 2028 and 1024 hidden neurons. |