I have simulated an industrial process by ASPEN and I want to optimize the operational parameters (decision) to maxmize the final yield (objective) while minmizing the energy comsumption (second objective)
Therefore, I would like to build a ML model to use it to optimize/decide the best operational parameters.
My plan is to use metaheuristic models such GA (Genetic Algorithm) but I have a difficulty to know what are the steps to build the model? How this optimization algorithm can be be implimented as a supervised ML Model?
Hi Mo Gamal
If I understand your scenario correctly, this is a multi-objective optimization problem. Please refer to the following links for some direction in going about solving your optimization problem:
1. https://medium.com/analytics-vidhya/optimization-modelling-in-python-multiple-objectives-760b9f1f26ee
2. https://stackoverflow.com/questions/16739118/genetic-algorithms-and-multi-objectives-optimization-on-python-libraries-tools
Evolutionary algorithms/ meta-heuristics (genetic algorithms, simulated annealing, etc) form a class of meta-heuristic search algorithms that solve for a solution to a particular combinatorial function or equation (They can also be applied to others as well, combinatronics are the most common application for evolutionary algorithms).
In contrast, machine learning algorithms create a model based on patterns that are found in the data. Both are not exactly the same if you are using it for optimization.
From what I understand from your question, you wish to build a model using macine learning based on the meta-heurstic algorithm of your choice. However, doing this is not optimization, but rather the modelling of your data from the evolutionary algorith,.
With that being said, a recent research area has come about that combines the qualities of evolutionary algorithms and deep learning techniques. Please search the term "Graph Neural Networks". From the research that I have read, it was used to find a solution to the Travelling Salesman Problem. However, I cannot recall if the solution was near-optimal or not.
Good luck!
Hi Shaheen,
Thanks for answering. Actually, I am still planning what is the best strategy to optimize the results. If I understood your comment correct, when using evolutionary algorithm with ML it will be like modelling the data. The question is, can optimization be done based on ML Alogirthms?, say using a supervised polynoimal algorithm to optimize the operational parameter to obtain the maxmium yield. Moreover, is it suitable when I have Multi-objectives?
Hi Mo Gamal
From my understanding, I do not think ML can be used for optimization. However, I could be wrong.
I did find a few resources that could be helpful, these are listed below
1. https://medium.com/swlh/ml-fundamentals-optimization-problems-and-how-to-solve-them-572c6ddf0a0b
2. https://www.datasciencecentral.com/an-easy-way-to-solve-complex-optimization-problems/
3. https://www.quora.com/Is-deep-learning-useful-for-optimization-problems-If-yes-how-does-it-compare-with-traditional-algorithms-such-as-simulated-annealing-evolutionary-algorithms-etc
I hope these help.
Good luck!
Dear MO,
Generally, machine learning methods have an algorithm to learn and generalise from recorded data to construct projections on new samples, and they do not use for optimising a design directly. There are a large number of optimisation methods that can be applied to your problem, but you need a fitness (objective) function in order to evaluate the new solution proposed by the optimisation method. You can easily apply an optimisation method if you have such a function.
Please let me know if you need more information in this way.
Here, the key questions would be: why would you want the ML model, and what is it a surrogate for?
If your simulator actually takes a bit of time (e.g. minutes to hours, or more) to evaluate a set of operational parameters, then you may want to try multi-objective Bayesian optimisers. For instance, here is an example application where we tried to perform design optimization using expensive computation fluid dynamic simulations.
Article Application of multi-objective Bayesian shape optimisation t...
On the other hand, if you would like to simply construct a data-driven (surrogate) model of the process and then identify the (near-)optimal set of operational parameters trading-off between multiple objectives, you may want to have a look at the following paper:
Article Data-driven multi-objective optimisation of coal-fired boile...
There are, of course, lots of work in the area of surrogate-assisted evolutionary optimisation. One key survey is as follows: Article Surrogate-assisted evolutionary computation: Recent advances...
- Badges
- Science topic
- Similar topics
- Linguistics
- Composition Studies
- Publication