Whether approving loans, supporting hiring decisions, prioritizing healthcare interventions, or assisting criminal justice systems, machine learning models are now operating in domains where fairness is not optional — it is essential.

Yet a persistent challenge remains:

How do we build models that are both accurate and fair?

For years, this has often been framed as a trade-off: improve fairness and risk losing predictive performance, or maximize accuracy and accept possible bias.

But what if that trade-off can be navigated more intelligently?

Our recent research, “Balancing Fairness and Accuracy Using Grammatical Evolution,” explores precisely that question through an unusual but powerful lens: evolutionary computation combined with causal reasoning.

Why Fairness in AI Is Still Hard

Modern machine learning systems often inherit historical and structural biases embedded in data. Even highly accurate models can generate systematically unequal outcomes for different groups.

This is particularly problematic in high-stakes settings.

Traditional fairness interventions often happen after a model is built:

• Pre-process the data
• Adjust the training objective
• Post-process predictions to satisfy fairness constraints

While valuable, these approaches often treat fairness as a corrective patch.

We wanted to explore a different idea:

What if fairness could be part of the search process itself?

Enter Grammatical Evolution

At the center of our work is Grammatical Evolution (GE) — a form of evolutionary computation that evolves structured solutions using grammar-based representations.

Rather than simply optimizing parameters, GE can evolve causal graph structures, enabling models that are:

• Predictive
• Fairness-aware
• Interpretable

That last property matters enormously.

In many fairness discussions, interpretability is often overlooked. But understanding why a model behaves fairly (or unfairly) is just as important as measuring the fairness outcome.

From Optimization to Fairness-Aware Evolution

Our work introduces three complementary approaches.

GE-α: Constrained Fairness Optimization

The first approach embeds fairness directly as a constraint during evolution.

Instead of evolving models solely for predictive accuracy, we ask:

Can evolution search for models that satisfy fairness criteria while remaining accurate?

Using Equalized Odds Difference as the fairness objective, GE-α explicitly searches for balanced solutions rather than treating fairness as an afterthought.

This turns fairness into part of the fitness landscape.

And that changes everything.

GE-β: A Minimax Perspective on Fairness

The second approach pushes further.

Rather than optimizing average behavior, GE-β uses a minimax formulation that focuses on worst-case trade-offs between error and unfairness.

This is important because responsible AI should not only perform well on average — it should be robust under difficult conditions.

The minimax perspective introduces a stronger notion of balance:

Not just “good enough” fairness.

But fairness that holds under pressure.

GE-γ: Causal Interventions for Explainable Fairness

This is perhaps the most exciting part of the work.

With GE-γ, we integrate causal reasoning through Average Causal Effect (ACE) and intervention-based node pruning.

The goal is not only to improve fairness metrics but to simplify and clarify the learned causal structures themselves.

Low-impact nodes can be removed.

Causal roles like confounders, mediators, and colliders are preserved.

The resulting models become more interpretable without sacrificing performance.

This opens an intriguing possibility:

Using evolution not only to optimize models, but to evolve explanations.

That is a very different vision of machine learning.

Why This Matters

What excites me most about this work is not just the fairness results.

It is the broader idea that evolutionary computation may be an underexplored engine for Responsible AI.

We often associate evolutionary algorithms with optimization benchmarks.

But they may also help us tackle problems involving:

• Fairness
• Transparency
• Robustness
• Causal structure discovery
• Multi-objective ethical trade-offs

That is a much richer role.

And perhaps a much more important one.

Beyond Accuracy-Centric AI

There is a quiet shift happening in AI research.

We are moving beyond asking:

“How accurate is the model?”

Toward asking:

• Is it fair?
• Is it explainable?
• Is it trustworthy?
• Can we intervene when harms emerge?

Those questions require new computational ideas.

We believe evolutionary methods have something valuable to contribute.

A Broader Vision

This work also points toward something larger:

A future where machine learning systems are not merely optimized for performance, but evolved under principles of responsibility.

Imagine search processes where objectives include:

• Accuracy
• Fairness
• Interpretability
• Causal validity
• Societal constraints

That begins to look less like conventional model training—

and more like engineering trustworthy intelligence.

Final Thoughts

Fairness and accuracy do not have to be opposing goals.

Sometimes the real opportunity lies in changing how we search for solutions altogether.

That is what makes evolutionary approaches so compelling.

They do not simply tune models.

They explore possibility spaces.

And in those spaces, we may discover better ways to build AI systems worthy of trust.

Our work on Balancing Fairness and Accuracy Using Grammatical Evolution is one small step in that direction.

I believe there is much more ahead.

If you work in fairness, causal machine learning, evolutionary computation, or Responsible AI, I’d love to hear your thoughts.

How should we evolve the next generation of trustworthy AI?

—
#ArtificialIntelligence #ResponsibleAI #FairnessInML #EvolutionaryComputation #CausalInference #ExplainableAI #MachineLearning #Research

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

GELAB – The Cutting Edge of Grammatical Evolution

The advent of cloud-based super-computing platforms has given rise to a Data Science (DS) boom. Many types of technological problems that were once considered prohibitively expensive to tackle are now candidates for exploration. Machine Learning (ML) tools that were valued only in academic environments are quickly being embraced by industrial giants and tiny startups alike.

Photo by exfordy

Analysis of Diversity in Grammatical Evolution – IEEE Conference Publication

Diversity is a much sought after aspect of any evolutionary system. More diversity means a cornucopia of diverse behaviors and traits among the individuals of a population. Lack of diversity, on the other hand, leads to a stagnant population whose individuals are more or less similar to each other.

Photo by nola.agent

GELAB and Hybrid Optimization Using Grammatical Evolution

Muhammad Adil Raja Aidan Murphy Conor Ryan Part of the Lecture Notes in Computer Science book series (LNCS, volume 12489) Grammatical Evolution (GE) is a well known technique for program synthesis and evolution. Much has been written in the past about its research and applications.

Following is the video presentation of the paper.

And here is the link to the git repository of GELAB.

adilraja/GELAB

GELAB: A Matlab Toolbox for Grammatical Evolution GitHub is home to over 50 million developers working together to host and review code, manage projects, and build software together. Millions of developers and companies build, ship, and maintain their software on GitHub – the largest and most advanced development platform in the world.

Photo by alexandre.lavrov

GitHub – adilraja/GELAB: GELAB: A Matlab Toolbox for Grammatical Evolution

GELAB: A Matlab Toolbox for Grammatical Evolution. Contribute to adilraja/GELAB development by creating an account on GitHub.

Photo by Barbara L. Hanson

Evolving MIMO multi-layered artificial neural networks using grammatical evolution

We’re upgrading the ACM DL, and would like your input. Please sign up to review new features, functionality and page designs.

Photo by stevendepolo

The presentation was about an application of association rule mining for predicting road accidents. Initially, I kept sitting dumb and idle. It was a bit hard to recall how association rule mining worked. After a while, as the presentation progressed, as a few people asked questions and especially as my own mind started generating questions about the work and association rule mining, I became a lot more engaged in the discourse.

A great conversation took place. Some people were actually doing the job of requesting others to stop asking too many questions. Too many questions definitely overwhelm the speaker. But isn’t it a good practice to face so many questions so as to develop experience in giving answers. Politicians are really adept at that. And this is possibly why they have great fortunes.

Given the current state of the art of the algorithm, I found it a bit clumsy. Well, it is a nice algorithm indeed in the sense that it figures out relationships within variables of a multi-dimensional dataset. But I think it could be improved.

My idea was to employ grammatical evolution to automate, upstage or drastically improve the algorithm. I personally think it could be good. To support my argument, I am sharing a related article that was published at a nice venue. Please give it a read. And yes, if you find this idea fascinating and want to work on it with me, please give me a shout. I think a nice FYP or even a postgraduate thesis could be drawn out from this.

An Intrusion-Detection Model Based on Fuzzy Class-Association-Rule Mining Using Genetic Network Programming – IEEE Journals & Magazine

As the Internet services spread all over the world, many kinds and a large number of security threats are increasing. Therefore, intrusion detection system

Photo by markhillary

GELAB – A Matlab Toolbox for Grammatical Evolution

In this paper, we present a Matlab version of libGE. libGE is a famous library for Grammatical Evolution (GE). GE was proposed initially in [1] as a tool for automatic programming. Ever since then,…

And here is a presentation about GELAB that is available on Youtube. I hope you like it.

There are quite a few projects I have in my mind about this wonderful tool. If you are willing to work on an interesting problem in machine learning, please get in touch.

Photo by O.S. Fisher

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