Our fusion reactor Miranda have dozens of parameters to adjust to obtain reactions.
Our simulator uses an easy algorithm to simulate our models running 16 threads but it is difficult to change all the parameters so we have few data to feed the learning algorithm, so we need a neural network with a very fast method that learns with few samples.
Dear Lopez;
Have you tried to resize your data? applying some transformations? Also you may want to preprocess your data to make the network training more efficient e.g. scaling between a specific range. Besides, you may also want to try different training functions e.g. TRAINLM is widely used because it is the fastest, but it requires significantly more memory...
Use Extreme Leaning Machine, no iteration. One step learning.
Simple description available:
Conference Paper Using ELM Techniques to Predict Data Centre VM Requests
Thanks to all,
Our mathematisian (@Daniel Lopez) will study at them(included RPROP or QuickProp).
It would help if there is a recommended C++ library of every one to include in our simulator
The basic idea is that is very difficult to change some parameters, as example the reactor dimensions, due an error would lead in grat building cost of new prototypes.
Perhaps fast algorithms is the key, like humans, the software must take decisions based on few data, but must use results to include further optimizations, perhaps using a fast algorithm as said Salam Ismael and then use a more powerful one involving array inverse as said Ibrahem.
In the actual system I use a file with inputs and outputs to train the network, but in the future I would like that the algorithm takes the reactor control that is extremely difficult due dozens of parameters must be chosen
Salam, the system can take some seconds to make some array operations. Does the ELM technique work well with 8-15 parameters? ¿does it taken in account second order derivatives?
For now, I am currently reading about Bayesian Neural Networks and it requires only one forward pass to get the correct output. Instead of optimizing NN parameters (weights), take the average overall weights by using Monte Carlo dropout to approximate the posterior distribution over network parameters p(W|X,Y). In this case, you are replacing the task of optimizing the NN parameters with the optimizing the parameters of a simple distribution like Gaussian. Of course, too many details are skipped here. I am referring to the paper
"What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision" by Alex Kendall and Yarin Gal both from university of Cambridge.
Article What Uncertainties Do We Need in Bayesian Deep Learning for ...
Well there are two meanings of my question:
1. Fastest algorithm that needs low computation power, that can be used in the realtime, as example during testing the reactor
2. Fast learning that needs low data. Here can be more compute intensive and include local and general array inverse calculations. It is used a compute intensive simulator that makes a lot of operations also, the IA will be used to reduce a lot the number of simulations. The results will be seen in the realtime
I think by redues no. of itrations, like use Bayesian neural network
- Badges
- Science topic