Modeling and tuning genetic algorithms

Here is our latest work that presents a solution on the problem of Service Chain Embedding. It is based on genetic algorithms and extends a previously published conference paper.

The interesting contribution of this paper is a modeling framework for the operation of genetic algorithms. Using this framework we prove that NP-hard problems are not computed efficiently by genetic algorithms and we define some properties for the problems that genetic algorithms compute efficiently. 

Another interesting contribution of this paper is a performance optimization mechanism for genetic algorithms which is also based on genetic computing. So you use one genetic algorithm in order to optimize the performance of another.

https://doi.org/10.1007/s10922-023-09771-y

Distributed Unsupervised Deep Learning

Our recently published paper, available here in open access mode, presents a deep learning method for network resource orchestration. There are a few features that make this method interesting.

It is build on a distributed multi-agent architecture.

It is based on Unsupervised Deep Learning, not a common feature for resource orchestration methods. The user essentially defines an objective and the agents try to accomplish it by training and then running deep neural networks and without further interaction with the user.

The agents share among them the most efficient models making the training process more efficient.

The neural networks are trained using genetic algorithms which is an innovative feature for unsupervised learning systems and speeds up the training procedure. It is actually interesting to use one system in order to train another system without explicitly describing the training process.

We are able to test this method by running simulations in large scale topologies. For this we have built an efficient network simulator resealed as an open source project.

https://doi.org/10.1109/ACCESS.2023.3308492