Artificial Intelligence systems are no longer confined to research labs. They influence decisions about loans, employment, healthcare, and criminal justice—domains where fairness is not optional. Yet, much of modern machine learning still treats fairness as a secondary concern: something to be fixed after a model has already learned its patterns.

One of the most persistent assumptions in this space is that fairness and accuracy are inherently at odds. Improve one, and the other must suffer. But is this trade-off inevitable—or is it simply a limitation of how we frame the problem?

In our recent work, A Multi-Objective Approach to Balance Fairness and Accuracy, we argue for a different perspective: fairness should be treated not as a constraint or post-processing correction, but as a first-class optimisation objective, explored alongside accuracy rather than subordinated to it. I would like to congratulate my co-authors about this who are: 1. the doctoral candidate Zahid Irfan, Dr. Roisin Loughran, and Professor Fergal Mc Caffery. Basically this is the work was done by Zahid, who is a colleague as well as a very close friend of mine.

Why Bias Persists in Machine Learning

Bias in AI systems often reflects deeper structural issues: biased data collection, historical inequalities, and spurious correlations that models eagerly exploit. When these correlations involve protected attributes—such as sex, age, or race—the resulting systems may achieve impressive accuracy while still producing unfair outcomes.

Traditional bias-mitigation approaches typically fall into three categories:

• Pre-processing, where the data is modified before training
• In-processing, where fairness is incorporated into the learning algorithm
• Post-processing, where predictions are adjusted after training

While all three have their place, many approaches operate largely as black-box fixes. They may improve a fairness metric, but often at the cost of interpretability and deeper understanding.

This is where causal modelling becomes essential.

Bringing Causality into the Picture

Correlation alone cannot tell us why a model behaves unfairly. Causal models, on the other hand, explicitly represent cause–effect relationships between variables.

We use Causal Bayesian Networks (CBNs)—directed acyclic graphs where nodes represent variables and edges encode causal influence. These structures allow us to reason about dependencies, confounders, and interventions, rather than relying solely on statistical association.

However, learning causal structures from data is a notoriously difficult problem. The search space of possible graphs grows exponentially, making exhaustive search infeasible.

To address this, we turned to Evolutionary Computation.

Evolving Causal Graphs with Grammatical Evolution

Our approach uses Grammatical Evolution (GE) to automatically generate and evolve causal graph structures. A context-free grammar constrains the search space to valid causal graphs, while still allowing a rich variety of structures to emerge.

Each individual in the evolutionary population represents a candidate causal graph. From this graph, we build a CBN, train it on data, and evaluate its performance.

Crucially, we do not evaluate performance using a single objective.

Fairness and Accuracy as Joint Objectives

Instead of collapsing everything into one score, we adopt a multi-objective optimisation framework using NSGA-II, a well-established evolutionary algorithm.

We optimise two objectives simultaneously:

1. Accuracy, measuring predictive performance
2. Fairness, measured using Equal Opportunity Difference (EOD), which captures disparities in true positive rates between protected groups

This produces not a single “best” model, but a Pareto front—a set of non-dominated solutions representing different fairness–accuracy trade-offs.

This is a powerful shift in mindset. Rather than asking “What is the best model?”, we ask:
“Which trade-off best fits the ethical and operational requirements of this domain?”

What We Observed

Using the German Credit dataset as a case study, our experiments showed that:

• It is possible to achieve very low fairness disparity while maintaining competitive accuracy
• Multiple causal graphs can yield similar performance, offering flexibility and interpretability
• The evolved graphs are non-trivial, capturing meaningful dependencies among features
• Practitioners can choose models that slightly sacrifice accuracy for substantial gains in fairness—or vice versa

Importantly, the causal graphs themselves provide insight. They allow us to inspect how features influence outcomes, opening the door to causal reasoning, domain validation, and future intervention analysis.

Why This Matters

Fair AI is not just about metrics—it’s about understanding.

By combining causality with multi-objective evolutionary optimisation, this work demonstrates that:

• Fairness does not have to be an afterthought
• Accuracy does not have to be blindly maximised
• Interpretability and performance can coexist

Most importantly, it reframes fairness as an optimisation problem, not a moral constraint imposed from outside the model.

Looking Ahead

Future directions include:

• Exploring additional fairness metrics to capture different notions of equity
• Extending experiments to larger and more diverse datasets
• Incorporating causal interventions and counterfactual analysis
• Further strengthening the link between ethical requirements and model design

As AI systems continue to shape society, approaches that integrate ethics, causality, and optimisation will be essential—not optional.

Fairness is not something we bolt onto AI.
It is something we design for

adilraja/libGEjava

libGEjava – A framework for Grammatical Evolution in Java. This is based on the original source code of GE in C++.

This version of libGE is simply a code-level translation of the original libGE that was written in C++. As a matter of fact, I initiated the translation in August, 2007. I even completed most of the translation at that time. However, for some good reasons, I had to abandon it. I reverted back to it a while ago and here we have a working piece of code. You can clone it and open it in NetBeans IDE. The idea now is to use it in a more productive way. I shall say more about it later on.

Photo by neiljs

The idea I have is to port this to GroovyLab. Groovy is a java-like scripting language. Moreover, all valid java code is also valid groovy code. Add to this the fact, that despite being a scripting language, and requiring an interpreter, Groovy code can also be pre-compiled and run on JRE.

Once GEVA is ported to GroovyLab, the resulting package can be run on Octave or Matlab. It is not necessary, however, to port the code to GroovyLab. GEVA can be wrapped into Groovy and pre-compiled. The code can then be called in Octave or Matlab. Having GE available in Matlab would mean that it would be easier to use it for research.

No Title

Grammatical evolution | IEEE Journals & Magazine | IEEE Xplore

Click to access tutorial.pdf

UCD NCRA – Software – GEVA

GEVA is no longer maintained. We keep the page below for archival reasons. PonyGE is the implementation of choice in our group. GEVA is an implementation of Grammatical Evolution in Java developed at UCD’s Natural Computing Research & Applications group. As well as providing the characteristic genotype-phenotype mapper of GE a search algorithm engine, and GUI are also provided.

Groovy

The latest innovations from dozens of Apache projects and their communities in a collaborative, vendor-neutral environment. There is a 2-day Groovy track covering latest news, testing and functional programming with Groovy, Spock, Grails, Micronaut, GORM, Gradle, DSLs, using Groovy for data science, with Apache Ignite, with Kubernetes, and more.

Photo by steve p2008

The key idea behind innovation in this feat is to have a realistic solar panel and to integrate it with a machine learning algorithm. The learning algorithm helps produce novel designs with trial and error, as is the case with any machine learning algorithm. I have been writing in the past about how to develop controllers for UAVs using machine learning. Similar ideas precisely map to this domain as well. The only difference being that instead of developing controllers, we evolve designs for solar panels.

I compiled a list of various online solar systems simulators. I hope that they can be used to come up with good innovative work on this.

Photo by Marufish

It is also not news that evolutionary algorithms are employed to create art. Several conferences and workshops are help each year with a section or chapter on evolutionary art. Evostar, for instance, hosts EvoMUSART every year. EvoMUSART has now grown into a full-fledged annual conference that is hosted as a part of the annual Evostar event. The conference is quite well respected.

Evostar 2015

Here is a glimpse of the abstracts of the papers that were nominated for best paper awards by the conference in 2015.

Evostar 2015

The page contains quite a lot of very nice themes both for music and visual arts. Although the explanations of some of the themes sound arcane, they are nonetheless interesting.

Addressing problems in music and art with the help of evolutionary algorithms and other biologically inspired themes does sound interesting indeed.

I would like to discuss an idea about creating artistic images using artificial intelligence techniques. I read about it quite a few years ago. However, I lost track of the source where I found it from. If I remember it correctly, it was called pixel art. It employs the famous ant colony optimization algorithm to convert a colour image to grey scale in a way that it appears to be sketched with a pencil to the user. So the work was all about doing this transformation. But the transformation did look quite brilliant. And the way the algorithm was applied to achieve the transformation was not short of brilliance in any way. Following is a similar approach.

Genetic and Evolutionary Computation – GECCO 2004

Following citation also looks good.

Animated drawings rendered by genetic programming | Proceedings of the 11th Annual conference on Genetic and evolutionary computation

Chakraborty, U. K. and Kang, H. W., Stroke-based rendering by evolutionary algorithm. India Annual Conference, 2004. Proceedings of the IEEE INDICON 2004. First, 2004. ]]Neufeld, C., Ross, B. and Ralph, W., The Evolution of Artistic Filters. The Art of Artificial Evolution, pages 335–356, 2008.

Photo by Song_sing

I contemplated a little on this idea today as I was asked by a couple of friends for ideas to work in cloud computing. So some of the ideas I gave in light of the memories I had from EuroGP are in order.

Evolutionary algorithms, and other machine learning algorithms, could be used to evolve (or improve) cloud operating systems, or parts of them. Implementation details aside, an evolutionary algorithm could silently monitor the user experience of a cloud operating system. One such operating system is OpenStack. Based on the feedback obtained by the monitors, the evolutionary algorithm could then update its future generations of solutions to those problems. The newer generations of solutions could be used as improvements to the existing solutions in the operating system.

For experimental purposes, evolutionary algorithms could be used to solve miniature problems that affect the user experience. As it may be envisioned, a lot of things add to the user experience of a complicated operating system for cloud computing. There is the graphical user interface that a user uses to interact and work with a computer or a cluster. Then there are parts of the kernel that should supposedly make the computing efficient and directly or indirectly affect the user experience.

A researcher could take one issue at a time. Let it be the graphical user interface, or a program that manages it. In the case of Ubuntu systems, it could be the unity desktop. Monitors could be embedded in a usability matrix which could keep track of various aspects of level of satisfaction of the user with the interface. Such monitors can then send data back to an evolutionary algorithm to update its state. Iteratively, the evolutionary algorithm can adapt and improve parts of the user interface (or unity desktop). Over multiple generations, and based on the feedback obtained from experience of many users, it can be hoped to improve the unity engine to great extent. Same can be done with other parts of the OS.

How to Master Ubuntu’s Unity Desktop: 8 Things You Need to Know

Ubuntu’s Unity desktop is a change of pace, whether you’re coming from Windows or another Linux distribution with a more traditional interface.

Photo by wwarby

In order to fully understand and appreciate the ideas presented in this article I strongly recommend you to read my two previous articles on this topic. These would help you understand the scope of these ideas. These titles are Simulated Drone Flying Competition and Explaining the Simulated Drone Flying Competition. You should also read these tutorials on genetic algorithms and especially on genetic programming. These tutorials would help you to understand these technologies and you would not find it much difficult to understand this tutorial. As a matter of fact you would not find it quite easy to understand this tutorial. All the ideas would come to your mind quite easily.

So let’s get started with this tutorial now. So the problem that I am discussing here is that how to evolve controllers for simulated drones. Let us begin with this basic function first that why do we need to evolve any controllers in the first place?

The answer to this question is simple. If you imagine how a plane is flown you would find it fairly natural to consider that the pilot who is responsible for flying the plane has to be able to control the plane somehow. As a matter of fact the guy sitting in the cockpit has access to a number of tools in his hands and under his feet through which he (or she) tries to steer the plane successfully to the desired destination during his moment by moment experience of flying. If all of those controllers were not there, or if he had not been using them to control the plane with his skills and experience of flying planes, he would simply not be able to fly the plane. The plane would simply crash resulting in tragedy.

This has possibly answered the question that why controllers are required to control a plane?

Why do we need to evolve controllers?

Consider that if you are trying to replace a human pilot in an aircraft with some sort of artificial intelligence that would fly the plane as well as a human being would. This can be a great idea. This is also a central theme behind designing drones. And in order to accomplish this task you would either have to develop a background in machine learning or artificial intelligence. And this also answers the question on as to why do we need to evolve controllers.

Now let us answer one more question: Why evolve controllers for simulated drones? The answer for this question is simple, although there could be quite a few reasons. And this is an extremely important question. The answer lies in the question that why do we need to evolve simulated drones in the first place? The reasons we would prefer to design drones in simulation lies in the expenditure it may require to test, try, design and evolve controllers for drones while employing real drones. Most of the machine learning algorithms employ hit and trial methods. We literally have to allow the algorithm to err while it tries to find optimum solutions. This is quite natural to suppose and understand as well that as new solutions are designed, or evolved, it is done so at the expense of bad solutions at times. And bad solutions and controllers can result in a lot of crashes, thus making employment of real drones for design of their controllers a very expensive expedition to undertake.

So as a result controllers for drones have to be designed in simulation. Whether or not the simulated controllers would be good enough for deployment in real drones depends partly on the quality of the controllers that have been designed and also on the ability of the simulation environment to mimic most types of real environments. If you want to design controllers that do other complex tasks besides ordinary flying, such as extinguishing fires or coordinate with other drones as they perform complex activities, you would have to develop simulation environments that can allow your drones to do exactly that.

How to evolve controllers for simulated drones then? This is our final question and in order to explain this, I would like to draw your attention to the tutorials about genetic algorithms and genetic programming. Both of them are population-based algorithms. The latter is a lot more powerful as it allows whole computer programs to be evolved. Both algorithms generate a huge population of individuals as they start. Then they evolve newer populations of individuals using genetic operators of crossover and mutation. They test each individual for its fitness to solve the underlying problem. In this problem a fitness score could be based on how well the set of controllers evolved allow the drone to perform the prescribed tasks of coordination while flying and carrying out the tasks. Once all the individuals of the population have been assigned fitness, a certain number of good individuals are kept and bad ones are littered. The good ones are used to make a new parent population of individuals. And a new evolutionary cycles begins.

At this stage it must be fairly intuitive for you to imagine for you that in the beginning the algorithm would generate a lot of bad and naive controllers. And they might result in a lot of crashes if real drones were employed. So we need nice simulation environments. It is only when a certain number of generations have elapsed, the search process may begin to find better individuals. And eventually, as we can hope, it would find an individual set of controllers that has all the dexterity of an adept human pilot in flying the drone.

The controller can be bench-marked at this stage and employed in real drones.

IEEE Xplore Document – Incremental evolution of autonomous controllers for unmanned aerial vehicles using multi-objective genetic programming

Autonomous navigation controllers were developed for fixed wing unmanned aerial vehicle (UAV) applications using incremental evolution with multi-objective

Advances in Unmanned Aerial Vehicles

Unmanned Aerial Vehicles (UAVs) have seen unprecedented levels of growth in military and civilian application domains. Fixed-wing aircraft, heavier or lighter than air, rotary-wing (rotorcraft, helicopters), vertical take-off and landing (VTOL) unmanned vehicles are being increasingly used in military and civilian domains for surveillance, reconnaissance, mapping, cartography, border patrol, inspection, homeland security, search and rescue, fire detection, agricultural imaging, traffic monitoring, to name just a few application domains.

Click to access MahboubiWang.pdf

Photo by Official U.S. Navy Imagery

I have been wanting to write a research proposal in this spirit for quite some time. And finally I wrote it and here it is. I hope that I would be able to evolve it to a better shape in some time.

This project is an implementation of the grammatical evolution (GE) framework. It is basically a hack of the original GE framework implemented in C++ by the BDS group of the university of Limerick. It is implemented in Java. It is named grammatical optimization because it is aimed at being more general. To this end, this means that it does not necessarily require a genetic algorithm as a search algorithm to be run on the back end. On the other hand, any nice search technique, such as particle swarm optimization can also be used. The source code of grammatical optimization can be found here.

null

null

Photo by seier+seier

Evolutionary nursery is a basic genetic algorithm implemented in java. It is suitable for simple numerical optimization tasks. Here you can find the source code of the project.

Photo by normanack