No-code ignorance

Lately, I see that platforms offering no-code AI have begun to appear. The issue with it is that we already have trouble understanding how complex AI models work and moving to no-code paradigms will just broaden our ignorance.

Generative AI and innovation

An interesting way to find out if the outcome of your research is innovative is to ask AI. Whenever I ask an LLM about my research findings and methods, I get the respond that my methods don’t follow some mainstream approach and I should be aligned with common practice. Then I know I am on the right track and I probably have developed something prototype or even innovative. 

AI currently cannot “think” out of the box and only reproduces known facts and of course hasn’t overcome yet human intelligence and we can sleep peacefully at night; for now.

Can we use AI to solve NP-Hard problems efficiently?

The growing research interest in Artificial Intelligence (AI) has led to the development of a plethora of methods and techniques that address difficult computational problems. The impressive progress on the field reasonably generates the question of whether we could use AI methods in order to solve efficiently traditionally hard problems or even to answer the NP vs P question.

These days the discussion on the possible relation of AI and NP-hardness revolves around two topics.

• Whether AI can resolve the NP vs P question
• If NP=P this will boost AI algorithms

On the first topic and at the time that this post is written, AI does not exhibit synthetic thinking so it is currently impossible to answer important problems as this. The second is really doubtful any way.

Let’s discuss a more practical aspect of the problem. AI methods deviate from traditional deterministic algorithms and in many cases overcome inherent obstacles of traditional computing. Then, could AI methods provide efficient solutions for NP-Hard problems?

There is huge literature on AI assisted approximation methods on NP-Hard problems that provide good and practical solutions for these problems sometimes achieving near optimal solutions for really hard instances of NP-Hard problems. So why not use them for actually solving them and provide exact solutions?

This reasonable question actually raises even more problems. Unlike traditional computing where Alan Turing solved the problem of modeling computation, AI methods lack general modeling of their operation. For some Machine Learning methods the researchers approach them experimentally and exhibit empirical evidences for their robustness without solid theoretical foundations.

The intense research in these issues gradually fills some of these gaps. In section 4 of our latest paper we provide a modeling framework for the operation of Genetic Algorithms (GA). Then we prove that GA may not solve efficiently classes of problems that include NP-Hard problems in every case; while they are pretty efficient on solving problems in P.

At least for now, it seems that AI can provide fast and practical approximate solutions for hard problems. But there is some pessimisms when we seek for exact solutions and confront NP-Hardness.

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