-One very promising direction is the use of Reinforcement Learning (RL). Where optimisation and control can be solved in a model-free or model -free rationale.
Have a look in the following for a recent overviewArticle Reinforcement Learning – Overview of Recent Progress and Imp...
Policy gradient methods is a very valuable tool of RL and it can be proven strong for PSE applications. For PSE applications, we have developed a technique for batch-to-batch optimisation:
Preprint Reinforcement Learning for Batch Bioprocess Optimization
ARTIFICIAL INTELLIGENCE - MACHINE LEARNING PROBLEM SOLVING ENVIRONMENT FOR COMPLEX DESIGN IN SCIENCE AND ENGINEERING
INTRODUCTION
Machine learning, a cornerstone of artificial and computational intelligence defined loosely as a means to build approximations from data, has played a key role in reducing the impact of the aforementioned challenge. The existing methods in the literature have been successfully applied in a host of practical domains spanning natural language processing, computer vision, recommender systems, etc. In complex design, machine learning is predominantly used for approximating the expensive physics-based simulations or multiagent-based simulations using supervised regression models, or more commonly, surrogate models. Despite the promising outcomes achieved thus far, it is deemed that the full capabilities of machine learning are yet to be fully unveiled and exploited in this domain.
Often, many calls to the analysis or simulation codes are often required to locate a near optimal solution. Optimization problems in which the evaluation of solutions is expensive arise in a variety of contexts. The reasons for the high cost of evaluation and their effect on how many design assessments can be afforded differ widely from one problem to another, as the following three examples may illustrate. (i) When evolving controllers for a simulated collective of robots, the fidelity of the physics simulator, the noise/stochasticity in the system, and the desire to obtain robots that are robust to rare events may all play a part in making simulation times very long. (ii) When evolving a novel protein for a specific binding target by synthesis of proteins in vitro and their subsequent screening, thousands of proteins may be synthesised in parallel but each further "generation" will take another 12 hours to process and will also have financial implications. (iii) When evolving a basic conceptual design for a new building, an architect evaluating the designs will suffer fatigue after several hours and will eventually have to stop.
To circumvent the abovementioned problems, some common practices have been investigated in this project: 1) the use of approximation models in lieu of the exact analysis code, and 2) parallelization of the analysis code evaluations. Approximation models are used to replace calls to the computationally expensive codes as often as possible in the evolutionary search process. These approximation models are commonly known as surrogate models or metamodels. Using approximation models, the computational burden can be greatly reduced since the efforts involved in building the surrogate model and optimization using it is much lower than the standard approach of directly coupling the simulation codes with the optimizer. Nevertheless, this approach does not always perform well when the approximation models are not managed properly. Inaccuracy of the models constructed is one of the many problems faced by most engineers and designers due to lack of data or curse of dimensionality. Hence, there is a need for methodologies to efficiently and effectively use approximation methods in optimization in the presence of such problems.
Further, it is noted that in current practices approximation models are typically built from scratch assuming zero prior knowledge, only relying on data sampled from the ongoing target problem of interest. However, it is contended that any practically useful intelligent system in an industrial setting will be faced with a large number of problems over a lifetime, with the problems likely sharing domain specific overlaps. With this in mind, we have also investigated three relatively advanced machine learning technologies that are especially developed to enable automatic knowledge transfer as a means of improving upon existing tabula rasa efforts. Our aim is to unveil meta-machine learning as a promising approach to enhance the efficiency of aircraft design and facilitate the realization of a more agile design process. Specifically, improving design procedures and methodologies so as to allow for rapid adaptation to change, lower development costs, and shorter time-to-market.
Approximation of Computational Expensive Analysis or Simulation Models
Aerodynamic Airfoil Wing Design
Discovery of Isomers in H2O(n) Using 1st Principal Methods
One of the well-known strength of stochastic optimization is also in the ability to partition the population of individuals among multiple computing nodes. Doing so allows sub-linear speedup in computation and even super-linear speedup if possible algorithmic speed-up is also considered. When applied to small scale dedicated and homogeneous computing nodes, this seems to be a very formidable solution. In real-life situation, there are many cases where heterogeneity exists, e.g. in a Grid/Cloud computing environment, which emphasizes on the seamless sharing of computing resources across laboratories and even geographical boundaries, heterogeneity of the resources in the sharing pool is inevitable. In addition to that, function evaluation time can vary in many cases, for instance, in the case where the objective function is a variable-fidelity function. In such situation a conventional parallelization without taking into account the heterogeneity of computing resources, might lead the EA to be ineffective. Hence, a suitable parallel optimization framework that fit in a heterogeneous computing environment while maintaining (or improving) the good search property of stochastic optimization is developed.
Parallel Hierarchical Genetic Algorithm on The Grid/Cloud
REFERENCES
W. M. Tan, Y. S. Ong, A. Gupta and C. K. Goh, "Multi-Problem Surrogates: Transfer Evolutionary Multiobjective Optimization of Computationally Expensive Problems", IEEE Transactions on Evolutionary Computation, In Press 2018. Available here: PDF file.
A. Gupta, Y. S. Ong and L. Feng, "Insights on Transfer Optimization: Because Experience is the Best Teacher", IEEE Transactions on Emerging Topics in Computational Intelligence, In Press 2017. Available here: PDF file.
H. Liu, J. F. Cai and Y. S. Ong, "Remarks on Multi-Output Gaussian Process Regression", Knowledge-Based Systems, In Press, 2018. Available here as PDF file.
H. Liu, Y. S. Ong and J. F. Cai, “A Survey of Adaptive Sampling for Global Metamodeling in support of Simulation-based Complex Engineering Design”, Structural and Multidisciplinary Optimization, In Press, 2017.
H. Liu, J. F. Cai and Y. S. Ong “An Adaptive Sampling Approach for Kriging Metamodeling by Maximizing Expected Prediction Error”, Computers and Chemical Engineering, In Press, 2017.
W. M. Tan, R. Sagarna, A. Gupta, Y. S. Ong, and C. K. Goh, , “Knowledge Transfer through Machine Learning in Aircraft Design”, IEEE Computational Intelligence Magazine, In Press, 2017, Available here as PDF file.
A. Kattan, A. Agapitos,Y. S. Ong, A. A. Alghamedi and M. O'Neill, “GP Made Faster with Semantic Surrogate Modelling”, Information Sciences, Vol. 355-356, pps. 169-185, 2016.
M. N. Le, Y. S. Ong, S. Menzel, Y. Jin and B. Sendhoff, "Evolution by Adapting Surrogates", Evolutionary Computation Journal, Accepted and In Press 2012.
S. D. Handoko, C. K. Kwoh and Y. S. Ong, "Feasibility Structure Modeling: An Effective Chaperon for Constrained Memetic Algorithms", IEEE Transactions on Evolutionary Computation, Accepted August 2009 and In Press. Available here as PDF file.
H. Soh, Y. S. Ong, Q. C. Nguyen, Q. H. Nguyen, M. S. Habibullah, T. Hung and J.-L. Kuo, “Discovering Unique, Low-Energy Pure Water Isomers: Memetic Exploration, Optimization and Landscape Analysis”, IEEE Transactions on Evolutionary Computation, Vol. 14, No. 3, pp. 419-437, Jun 2010. Available here as PDF file.
D. Lim, Y. Jin, Y. S. Ong and B. Sendhoff, "Generalizing Surrogate-assisted Evolutionary Computation", IEEE Transactions on Evolutionary Computation, Vol. 14, No. 3, pp. 329-355, Jun 2010. Available here as PDF file. *Source code Download*.
Q. C. Nguyen, Y. S. Ong, H. Soh and J.-L. Kuo, "Multiscale Approach to Explore the Potential Energy Surface of Water Clusters (H2O)8 n
-One very promising direction is the use of Reinforcement Learning (RL). Where optimisation and control can be solved in a model-free or model -free rationale.
Have a look in the following for a recent overviewArticle Reinforcement Learning – Overview of Recent Progress and Imp...
Policy gradient methods is a very valuable tool of RL and it can be proven strong for PSE applications. For PSE applications, we have developed a technique for batch-to-batch optimisation:
Preprint Reinforcement Learning for Batch Bioprocess Optimization
Different from the continuous and reduced space for Feature Selection found in RR, RF, LASSO, LAR, ENR, ANN/DL, RBF, SVM, KNN, etc., we developed a methodology that automatically searches for a discrete subset of the original features, basis functions or super-surrogate / -proxy / -correlation algebraic model terms in full-space to enable engineering-based constraint handling – also known as subset selection (see: https://aiche.confex.com/aiche/2019/meetingapp.cgi/Paper/565801)
The methodology combines theory-driven (engineering-based) and data-driven (empirical-based) aspects in an "MIP-Based" Machine Learning based on discrete subset or feature selection (MILP / MIQP) for algebraic and optimizable sub-models. Our work is primarily focused on providing better proxy or surrogate sub-models for Planning, Scheduling, Coordinating (Plant-Wide RTO) and single-unit RTO as an optimization and control for both profit and performance improvements. We use existing process engineering simulators to generate unlimited synthetic data free of measurement noise and most importantly feedback (inputs as functions of outputs) is essential to identifying the structure / shape of these proxy sub-models. Then, we use historical routine operating data and well-designed plant trials to estimate the parameters and provide observability and variance estimates for better sense-making.